mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 1 | /* |
fbarchard@google.com | b0c9797 | 2012-08-08 19:04:24 +0000 | [diff] [blame] | 2 | * Copyright 2011 The LibYuv Project Authors. All rights reserved. |
mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license |
| 5 | * that can be found in the LICENSE file in the root of the source |
| 6 | * tree. An additional intellectual property rights grant can be found |
| 7 | * in the file PATENTS. All contributing project authors may |
| 8 | * be found in the AUTHORS file in the root of the source tree. |
| 9 | */ |
| 10 | |
fbarchard@google.com | 142f6c4 | 2012-09-18 20:56:51 +0000 | [diff] [blame] | 11 | #include "libyuv/row.h" |
mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 12 | |
frkoenig@google.com | c82af4a | 2011-11-11 00:54:34 +0000 | [diff] [blame] | 13 | #include "libyuv/basic_types.h" |
| 14 | |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 15 | #ifdef __cplusplus |
| 16 | namespace libyuv { |
mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 17 | extern "C" { |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 18 | #endif |
mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 19 | |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 20 | // This module is for GCC x86 and x64 |
fbarchard@google.com | d2f4413 | 2012-04-04 21:53:27 +0000 | [diff] [blame] | 21 | #if !defined(YUV_DISABLE_ASM) && (defined(__x86_64__) || defined(__i386__)) |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 22 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 23 | // GCC 4.2 on OSX has link error when passing static or const to inline. |
| 24 | // TODO(fbarchard): Use static const when gcc 4.2 support is dropped. |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 25 | #ifdef __APPLE__ |
| 26 | #define CONST |
| 27 | #else |
| 28 | #define CONST static const |
| 29 | #endif |
| 30 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 31 | #ifdef HAS_ARGBTOYROW_SSSE3 |
| 32 | |
| 33 | // Constants for ARGB |
| 34 | CONST vec8 kARGBToY = { |
| 35 | 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0 |
| 36 | }; |
| 37 | |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 38 | CONST vec8 kARGBToU = { |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 39 | 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0 |
| 40 | }; |
| 41 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 42 | CONST vec8 kARGBToV = { |
| 43 | -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, |
| 44 | }; |
| 45 | |
| 46 | // Constants for BGRA |
| 47 | CONST vec8 kBGRAToY = { |
| 48 | 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13 |
| 49 | }; |
| 50 | |
| 51 | CONST vec8 kBGRAToU = { |
| 52 | 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112 |
| 53 | }; |
| 54 | |
| 55 | CONST vec8 kBGRAToV = { |
| 56 | 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18 |
| 57 | }; |
| 58 | |
| 59 | // Constants for ABGR |
| 60 | CONST vec8 kABGRToY = { |
| 61 | 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0 |
| 62 | }; |
| 63 | |
| 64 | CONST vec8 kABGRToU = { |
| 65 | -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0 |
| 66 | }; |
| 67 | |
| 68 | CONST vec8 kABGRToV = { |
| 69 | 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0 |
| 70 | }; |
| 71 | |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 72 | // Constants for RGBA. |
| 73 | CONST vec8 kRGBAToY = { |
| 74 | 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33 |
| 75 | }; |
| 76 | |
| 77 | CONST vec8 kRGBAToU = { |
| 78 | 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38 |
| 79 | }; |
| 80 | |
| 81 | CONST vec8 kRGBAToV = { |
| 82 | 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112 |
| 83 | }; |
| 84 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 85 | CONST uvec8 kAddY16 = { |
| 86 | 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 87 | }; |
fbarchard@google.com | 2430e04 | 2011-11-11 21:57:06 +0000 | [diff] [blame] | 88 | |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 89 | CONST uvec8 kAddUV128 = { |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 90 | 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, |
| 91 | 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u |
| 92 | }; |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 93 | |
fbarchard@google.com | ba1f526 | 2012-01-12 19:22:41 +0000 | [diff] [blame] | 94 | // Shuffle table for converting RGB24 to ARGB. |
| 95 | CONST uvec8 kShuffleMaskRGB24ToARGB = { |
fbarchard@google.com | 9394ed9 | 2011-10-31 21:36:47 +0000 | [diff] [blame] | 96 | 0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u |
| 97 | }; |
| 98 | |
| 99 | // Shuffle table for converting RAW to ARGB. |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 100 | CONST uvec8 kShuffleMaskRAWToARGB = { |
fbarchard@google.com | 9394ed9 | 2011-10-31 21:36:47 +0000 | [diff] [blame] | 101 | 2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u |
| 102 | }; |
| 103 | |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 104 | // Shuffle table for converting ABGR to ARGB. |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 105 | CONST uvec8 kShuffleMaskABGRToARGB = { |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 106 | 2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u |
| 107 | }; |
| 108 | |
| 109 | // Shuffle table for converting BGRA to ARGB. |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 110 | CONST uvec8 kShuffleMaskBGRAToARGB = { |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 111 | 3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u |
| 112 | }; |
| 113 | |
fbarchard@google.com | b8eabfe | 2012-09-14 06:59:31 +0000 | [diff] [blame] | 114 | // Shuffle table for converting RGBA to ARGB. |
| 115 | CONST uvec8 kShuffleMaskRGBAToARGB = { |
| 116 | 1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u |
| 117 | }; |
| 118 | |
fbarchard@google.com | a2cc341 | 2012-09-14 18:09:41 +0000 | [diff] [blame] | 119 | // Shuffle table for converting ARGB to RGBA. |
| 120 | CONST uvec8 kShuffleMaskARGBToRGBA = { |
| 121 | 3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u |
| 122 | }; |
| 123 | |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 124 | // Shuffle table for converting ARGB to RGB24. |
| 125 | CONST uvec8 kShuffleMaskARGBToRGB24 = { |
| 126 | 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u |
| 127 | }; |
| 128 | |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 129 | // Shuffle table for converting ARGB to RAW. |
| 130 | CONST uvec8 kShuffleMaskARGBToRAW = { |
fbarchard@google.com | f3fb7b6 | 2012-03-29 23:19:34 +0000 | [diff] [blame] | 131 | 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 132 | }; |
| 133 | |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 134 | // Shuffle table for converting ARGBToRGB24 for I422ToRGB24. First 8 + next 4 |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 135 | CONST uvec8 kShuffleMaskARGBToRGB24_0 = { |
| 136 | 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u |
| 137 | }; |
| 138 | |
| 139 | // Shuffle table for converting ARGB to RAW. |
| 140 | CONST uvec8 kShuffleMaskARGBToRAW_0 = { |
| 141 | 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u |
| 142 | }; |
| 143 | |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 144 | void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 145 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 146 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 147 | "pslld $0x18,%%xmm5 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 148 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 149 | "1: \n" |
| 150 | "movq (%0),%%xmm0 \n" |
| 151 | "lea 0x8(%0),%0 \n" |
| 152 | "punpcklbw %%xmm0,%%xmm0 \n" |
| 153 | "movdqa %%xmm0,%%xmm1 \n" |
| 154 | "punpcklwd %%xmm0,%%xmm0 \n" |
| 155 | "punpckhwd %%xmm1,%%xmm1 \n" |
| 156 | "por %%xmm5,%%xmm0 \n" |
| 157 | "por %%xmm5,%%xmm1 \n" |
| 158 | "movdqa %%xmm0,(%1) \n" |
| 159 | "movdqa %%xmm1,0x10(%1) \n" |
| 160 | "lea 0x20(%1),%1 \n" |
| 161 | "sub $0x8,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 162 | "jg 1b \n" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 163 | : "+r"(src_y), // %0 |
| 164 | "+r"(dst_argb), // %1 |
| 165 | "+r"(pix) // %2 |
| 166 | : |
| 167 | : "memory", "cc" |
| 168 | #if defined(__SSE2__) |
| 169 | , "xmm0", "xmm1", "xmm5" |
| 170 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 171 | ); |
fbarchard@google.com | 585a126 | 2011-10-28 23:51:08 +0000 | [diff] [blame] | 172 | } |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 173 | |
fbarchard@google.com | 00b69a2 | 2012-11-02 06:03:28 +0000 | [diff] [blame] | 174 | void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb, |
| 175 | int pix) { |
| 176 | asm volatile ( |
| 177 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 178 | "pslld $0x18,%%xmm5 \n" |
| 179 | ".p2align 4 \n" |
| 180 | "1: \n" |
| 181 | "movq (%0),%%xmm0 \n" |
| 182 | "lea 0x8(%0),%0 \n" |
| 183 | "punpcklbw %%xmm0,%%xmm0 \n" |
| 184 | "movdqa %%xmm0,%%xmm1 \n" |
| 185 | "punpcklwd %%xmm0,%%xmm0 \n" |
| 186 | "punpckhwd %%xmm1,%%xmm1 \n" |
| 187 | "por %%xmm5,%%xmm0 \n" |
| 188 | "por %%xmm5,%%xmm1 \n" |
| 189 | "movdqu %%xmm0,(%1) \n" |
| 190 | "movdqu %%xmm1,0x10(%1) \n" |
| 191 | "lea 0x20(%1),%1 \n" |
| 192 | "sub $0x8,%2 \n" |
| 193 | "jg 1b \n" |
| 194 | : "+r"(src_y), // %0 |
| 195 | "+r"(dst_argb), // %1 |
| 196 | "+r"(pix) // %2 |
| 197 | : |
| 198 | : "memory", "cc" |
| 199 | #if defined(__SSE2__) |
| 200 | , "xmm0", "xmm1", "xmm5" |
| 201 | #endif |
| 202 | ); |
| 203 | } |
| 204 | |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 205 | void ABGRToARGBRow_SSSE3(const uint8* src_abgr, uint8* dst_argb, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 206 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 207 | "movdqa %3,%%xmm5 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 208 | "sub %0,%1 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 209 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 210 | "1: \n" |
| 211 | "movdqa (%0),%%xmm0 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 212 | "pshufb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 213 | "sub $0x4,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 214 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 215 | "lea 0x10(%0),%0 \n" |
| 216 | "jg 1b \n" |
| 217 | |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 218 | : "+r"(src_abgr), // %0 |
| 219 | "+r"(dst_argb), // %1 |
| 220 | "+r"(pix) // %2 |
| 221 | : "m"(kShuffleMaskABGRToARGB) // %3 |
| 222 | : "memory", "cc" |
| 223 | #if defined(__SSE2__) |
| 224 | , "xmm0", "xmm5" |
fbarchard@google.com | 585a126 | 2011-10-28 23:51:08 +0000 | [diff] [blame] | 225 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 226 | ); |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 227 | } |
| 228 | |
| 229 | void BGRAToARGBRow_SSSE3(const uint8* src_bgra, uint8* dst_argb, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 230 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 231 | "movdqa %3,%%xmm5 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 232 | "sub %0,%1 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 233 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 234 | "1: \n" |
| 235 | "movdqa (%0),%%xmm0 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 236 | "pshufb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 237 | "sub $0x4,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 238 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 239 | "lea 0x10(%0),%0 \n" |
| 240 | "jg 1b \n" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 241 | : "+r"(src_bgra), // %0 |
| 242 | "+r"(dst_argb), // %1 |
| 243 | "+r"(pix) // %2 |
| 244 | : "m"(kShuffleMaskBGRAToARGB) // %3 |
| 245 | : "memory", "cc" |
| 246 | #if defined(__SSE2__) |
| 247 | , "xmm0", "xmm5" |
| 248 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 249 | ); |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 250 | } |
| 251 | |
fbarchard@google.com | b8eabfe | 2012-09-14 06:59:31 +0000 | [diff] [blame] | 252 | void RGBAToARGBRow_SSSE3(const uint8* src_rgba, uint8* dst_argb, int pix) { |
| 253 | asm volatile ( |
| 254 | "movdqa %3,%%xmm5 \n" |
| 255 | "sub %0,%1 \n" |
| 256 | ".p2align 4 \n" |
| 257 | "1: \n" |
| 258 | "movdqa (%0),%%xmm0 \n" |
| 259 | "pshufb %%xmm5,%%xmm0 \n" |
| 260 | "sub $0x4,%2 \n" |
| 261 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 262 | "lea 0x10(%0),%0 \n" |
| 263 | "jg 1b \n" |
| 264 | |
| 265 | : "+r"(src_rgba), // %0 |
| 266 | "+r"(dst_argb), // %1 |
| 267 | "+r"(pix) // %2 |
| 268 | : "m"(kShuffleMaskRGBAToARGB) // %3 |
| 269 | : "memory", "cc" |
| 270 | #if defined(__SSE2__) |
| 271 | , "xmm0", "xmm5" |
| 272 | #endif |
| 273 | ); |
| 274 | } |
| 275 | |
fbarchard@google.com | a2cc341 | 2012-09-14 18:09:41 +0000 | [diff] [blame] | 276 | void ARGBToRGBARow_SSSE3(const uint8* src_argb, uint8* dst_rgba, int pix) { |
| 277 | asm volatile ( |
| 278 | "movdqa %3,%%xmm5 \n" |
| 279 | "sub %0,%1 \n" |
| 280 | ".p2align 4 \n" |
| 281 | "1: \n" |
| 282 | "movdqa (%0),%%xmm0 \n" |
| 283 | "pshufb %%xmm5,%%xmm0 \n" |
| 284 | "sub $0x4,%2 \n" |
| 285 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 286 | "lea 0x10(%0),%0 \n" |
| 287 | "jg 1b \n" |
| 288 | |
| 289 | : "+r"(src_argb), // %0 |
| 290 | "+r"(dst_rgba), // %1 |
| 291 | "+r"(pix) // %2 |
| 292 | : "m"(kShuffleMaskARGBToRGBA) // %3 |
| 293 | : "memory", "cc" |
| 294 | #if defined(__SSE2__) |
| 295 | , "xmm0", "xmm5" |
| 296 | #endif |
| 297 | ); |
| 298 | } |
| 299 | |
fbarchard@google.com | ba1f526 | 2012-01-12 19:22:41 +0000 | [diff] [blame] | 300 | void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 301 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 302 | "pcmpeqb %%xmm5,%%xmm5 \n" // generate mask 0xff000000 |
| 303 | "pslld $0x18,%%xmm5 \n" |
| 304 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 305 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 306 | "1: \n" |
| 307 | "movdqu (%0),%%xmm0 \n" |
| 308 | "movdqu 0x10(%0),%%xmm1 \n" |
| 309 | "movdqu 0x20(%0),%%xmm3 \n" |
| 310 | "lea 0x30(%0),%0 \n" |
| 311 | "movdqa %%xmm3,%%xmm2 \n" |
| 312 | "palignr $0x8,%%xmm1,%%xmm2 \n" |
| 313 | "pshufb %%xmm4,%%xmm2 \n" |
| 314 | "por %%xmm5,%%xmm2 \n" |
| 315 | "palignr $0xc,%%xmm0,%%xmm1 \n" |
| 316 | "pshufb %%xmm4,%%xmm0 \n" |
| 317 | "movdqa %%xmm2,0x20(%1) \n" |
| 318 | "por %%xmm5,%%xmm0 \n" |
| 319 | "pshufb %%xmm4,%%xmm1 \n" |
| 320 | "movdqa %%xmm0,(%1) \n" |
| 321 | "por %%xmm5,%%xmm1 \n" |
| 322 | "palignr $0x4,%%xmm3,%%xmm3 \n" |
| 323 | "pshufb %%xmm4,%%xmm3 \n" |
| 324 | "movdqa %%xmm1,0x10(%1) \n" |
| 325 | "por %%xmm5,%%xmm3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 326 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 327 | "movdqa %%xmm3,0x30(%1) \n" |
| 328 | "lea 0x40(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 329 | "jg 1b \n" |
fbarchard@google.com | ba1f526 | 2012-01-12 19:22:41 +0000 | [diff] [blame] | 330 | : "+r"(src_rgb24), // %0 |
fbarchard@google.com | 9394ed9 | 2011-10-31 21:36:47 +0000 | [diff] [blame] | 331 | "+r"(dst_argb), // %1 |
| 332 | "+r"(pix) // %2 |
fbarchard@google.com | ba1f526 | 2012-01-12 19:22:41 +0000 | [diff] [blame] | 333 | : "m"(kShuffleMaskRGB24ToARGB) // %3 |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 334 | : "memory", "cc" |
| 335 | #if defined(__SSE2__) |
| 336 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 337 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 338 | ); |
fbarchard@google.com | 585a126 | 2011-10-28 23:51:08 +0000 | [diff] [blame] | 339 | } |
| 340 | |
fbarchard@google.com | 9394ed9 | 2011-10-31 21:36:47 +0000 | [diff] [blame] | 341 | void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 342 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 343 | "pcmpeqb %%xmm5,%%xmm5 \n" // generate mask 0xff000000 |
| 344 | "pslld $0x18,%%xmm5 \n" |
| 345 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 346 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 347 | "1: \n" |
| 348 | "movdqu (%0),%%xmm0 \n" |
| 349 | "movdqu 0x10(%0),%%xmm1 \n" |
| 350 | "movdqu 0x20(%0),%%xmm3 \n" |
| 351 | "lea 0x30(%0),%0 \n" |
| 352 | "movdqa %%xmm3,%%xmm2 \n" |
| 353 | "palignr $0x8,%%xmm1,%%xmm2 \n" |
| 354 | "pshufb %%xmm4,%%xmm2 \n" |
| 355 | "por %%xmm5,%%xmm2 \n" |
| 356 | "palignr $0xc,%%xmm0,%%xmm1 \n" |
| 357 | "pshufb %%xmm4,%%xmm0 \n" |
| 358 | "movdqa %%xmm2,0x20(%1) \n" |
| 359 | "por %%xmm5,%%xmm0 \n" |
| 360 | "pshufb %%xmm4,%%xmm1 \n" |
| 361 | "movdqa %%xmm0,(%1) \n" |
| 362 | "por %%xmm5,%%xmm1 \n" |
| 363 | "palignr $0x4,%%xmm3,%%xmm3 \n" |
| 364 | "pshufb %%xmm4,%%xmm3 \n" |
| 365 | "movdqa %%xmm1,0x10(%1) \n" |
| 366 | "por %%xmm5,%%xmm3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 367 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 368 | "movdqa %%xmm3,0x30(%1) \n" |
| 369 | "lea 0x40(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 370 | "jg 1b \n" |
fbarchard@google.com | 9394ed9 | 2011-10-31 21:36:47 +0000 | [diff] [blame] | 371 | : "+r"(src_raw), // %0 |
| 372 | "+r"(dst_argb), // %1 |
| 373 | "+r"(pix) // %2 |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 374 | : "m"(kShuffleMaskRAWToARGB) // %3 |
| 375 | : "memory", "cc" |
| 376 | #if defined(__SSE2__) |
| 377 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 378 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 379 | ); |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 380 | } |
| 381 | |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 382 | void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 383 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 384 | "mov $0x1080108,%%eax \n" |
| 385 | "movd %%eax,%%xmm5 \n" |
| 386 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
fbarchard@google.com | 6d6b770 | 2012-06-04 15:29:15 +0000 | [diff] [blame] | 387 | "mov $0x20802080,%%eax \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 388 | "movd %%eax,%%xmm6 \n" |
| 389 | "pshufd $0x0,%%xmm6,%%xmm6 \n" |
| 390 | "pcmpeqb %%xmm3,%%xmm3 \n" |
| 391 | "psllw $0xb,%%xmm3 \n" |
| 392 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 393 | "psllw $0xa,%%xmm4 \n" |
| 394 | "psrlw $0x5,%%xmm4 \n" |
| 395 | "pcmpeqb %%xmm7,%%xmm7 \n" |
| 396 | "psllw $0x8,%%xmm7 \n" |
| 397 | "sub %0,%1 \n" |
| 398 | "sub %0,%1 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 399 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 400 | "1: \n" |
| 401 | "movdqu (%0),%%xmm0 \n" |
| 402 | "movdqa %%xmm0,%%xmm1 \n" |
| 403 | "movdqa %%xmm0,%%xmm2 \n" |
| 404 | "pand %%xmm3,%%xmm1 \n" |
| 405 | "psllw $0xb,%%xmm2 \n" |
| 406 | "pmulhuw %%xmm5,%%xmm1 \n" |
| 407 | "pmulhuw %%xmm5,%%xmm2 \n" |
| 408 | "psllw $0x8,%%xmm1 \n" |
| 409 | "por %%xmm2,%%xmm1 \n" |
| 410 | "pand %%xmm4,%%xmm0 \n" |
| 411 | "pmulhuw %%xmm6,%%xmm0 \n" |
| 412 | "por %%xmm7,%%xmm0 \n" |
| 413 | "movdqa %%xmm1,%%xmm2 \n" |
| 414 | "punpcklbw %%xmm0,%%xmm1 \n" |
| 415 | "punpckhbw %%xmm0,%%xmm2 \n" |
| 416 | "movdqa %%xmm1,(%1,%0,2) \n" |
| 417 | "movdqa %%xmm2,0x10(%1,%0,2) \n" |
| 418 | "lea 0x10(%0),%0 \n" |
| 419 | "sub $0x8,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 420 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 421 | : "+r"(src), // %0 |
| 422 | "+r"(dst), // %1 |
| 423 | "+r"(pix) // %2 |
| 424 | : |
| 425 | : "memory", "cc", "eax" |
| 426 | #if defined(__SSE2__) |
| 427 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
| 428 | #endif |
| 429 | ); |
| 430 | } |
| 431 | |
| 432 | void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 433 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 434 | "mov $0x1080108,%%eax \n" |
| 435 | "movd %%eax,%%xmm5 \n" |
| 436 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
| 437 | "mov $0x42004200,%%eax \n" |
| 438 | "movd %%eax,%%xmm6 \n" |
| 439 | "pshufd $0x0,%%xmm6,%%xmm6 \n" |
| 440 | "pcmpeqb %%xmm3,%%xmm3 \n" |
| 441 | "psllw $0xb,%%xmm3 \n" |
| 442 | "movdqa %%xmm3,%%xmm4 \n" |
| 443 | "psrlw $0x6,%%xmm4 \n" |
| 444 | "pcmpeqb %%xmm7,%%xmm7 \n" |
| 445 | "psllw $0x8,%%xmm7 \n" |
| 446 | "sub %0,%1 \n" |
| 447 | "sub %0,%1 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 448 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 449 | "1: \n" |
| 450 | "movdqu (%0),%%xmm0 \n" |
| 451 | "movdqa %%xmm0,%%xmm1 \n" |
| 452 | "movdqa %%xmm0,%%xmm2 \n" |
| 453 | "psllw $0x1,%%xmm1 \n" |
| 454 | "psllw $0xb,%%xmm2 \n" |
| 455 | "pand %%xmm3,%%xmm1 \n" |
| 456 | "pmulhuw %%xmm5,%%xmm2 \n" |
| 457 | "pmulhuw %%xmm5,%%xmm1 \n" |
| 458 | "psllw $0x8,%%xmm1 \n" |
| 459 | "por %%xmm2,%%xmm1 \n" |
| 460 | "movdqa %%xmm0,%%xmm2 \n" |
| 461 | "pand %%xmm4,%%xmm0 \n" |
| 462 | "psraw $0x8,%%xmm2 \n" |
| 463 | "pmulhuw %%xmm6,%%xmm0 \n" |
| 464 | "pand %%xmm7,%%xmm2 \n" |
| 465 | "por %%xmm2,%%xmm0 \n" |
| 466 | "movdqa %%xmm1,%%xmm2 \n" |
| 467 | "punpcklbw %%xmm0,%%xmm1 \n" |
| 468 | "punpckhbw %%xmm0,%%xmm2 \n" |
| 469 | "movdqa %%xmm1,(%1,%0,2) \n" |
| 470 | "movdqa %%xmm2,0x10(%1,%0,2) \n" |
| 471 | "lea 0x10(%0),%0 \n" |
| 472 | "sub $0x8,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 473 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 474 | : "+r"(src), // %0 |
| 475 | "+r"(dst), // %1 |
| 476 | "+r"(pix) // %2 |
| 477 | : |
| 478 | : "memory", "cc", "eax" |
| 479 | #if defined(__SSE2__) |
| 480 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
| 481 | #endif |
| 482 | ); |
| 483 | } |
| 484 | |
| 485 | void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 486 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 487 | "mov $0xf0f0f0f,%%eax \n" |
| 488 | "movd %%eax,%%xmm4 \n" |
| 489 | "pshufd $0x0,%%xmm4,%%xmm4 \n" |
| 490 | "movdqa %%xmm4,%%xmm5 \n" |
| 491 | "pslld $0x4,%%xmm5 \n" |
| 492 | "sub %0,%1 \n" |
| 493 | "sub %0,%1 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 494 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 495 | "1: \n" |
| 496 | "movdqu (%0),%%xmm0 \n" |
| 497 | "movdqa %%xmm0,%%xmm2 \n" |
| 498 | "pand %%xmm4,%%xmm0 \n" |
| 499 | "pand %%xmm5,%%xmm2 \n" |
| 500 | "movdqa %%xmm0,%%xmm1 \n" |
| 501 | "movdqa %%xmm2,%%xmm3 \n" |
| 502 | "psllw $0x4,%%xmm1 \n" |
| 503 | "psrlw $0x4,%%xmm3 \n" |
| 504 | "por %%xmm1,%%xmm0 \n" |
| 505 | "por %%xmm3,%%xmm2 \n" |
| 506 | "movdqa %%xmm0,%%xmm1 \n" |
| 507 | "punpcklbw %%xmm2,%%xmm0 \n" |
| 508 | "punpckhbw %%xmm2,%%xmm1 \n" |
| 509 | "movdqa %%xmm0,(%1,%0,2) \n" |
| 510 | "movdqa %%xmm1,0x10(%1,%0,2) \n" |
| 511 | "lea 0x10(%0),%0 \n" |
| 512 | "sub $0x8,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 513 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 514 | : "+r"(src), // %0 |
| 515 | "+r"(dst), // %1 |
| 516 | "+r"(pix) // %2 |
| 517 | : |
| 518 | : "memory", "cc", "eax" |
| 519 | #if defined(__SSE2__) |
| 520 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 521 | #endif |
| 522 | ); |
| 523 | } |
| 524 | |
| 525 | void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 526 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 527 | "movdqa %3,%%xmm6 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 528 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 529 | "1: \n" |
| 530 | "movdqa (%0),%%xmm0 \n" |
| 531 | "movdqa 0x10(%0),%%xmm1 \n" |
| 532 | "movdqa 0x20(%0),%%xmm2 \n" |
| 533 | "movdqa 0x30(%0),%%xmm3 \n" |
| 534 | "lea 0x40(%0),%0 \n" |
| 535 | "pshufb %%xmm6,%%xmm0 \n" |
| 536 | "pshufb %%xmm6,%%xmm1 \n" |
| 537 | "pshufb %%xmm6,%%xmm2 \n" |
| 538 | "pshufb %%xmm6,%%xmm3 \n" |
| 539 | "movdqa %%xmm1,%%xmm4 \n" |
| 540 | "psrldq $0x4,%%xmm1 \n" |
| 541 | "pslldq $0xc,%%xmm4 \n" |
| 542 | "movdqa %%xmm2,%%xmm5 \n" |
| 543 | "por %%xmm4,%%xmm0 \n" |
| 544 | "pslldq $0x8,%%xmm5 \n" |
| 545 | "movdqa %%xmm0,(%1) \n" |
| 546 | "por %%xmm5,%%xmm1 \n" |
| 547 | "psrldq $0x8,%%xmm2 \n" |
| 548 | "pslldq $0x4,%%xmm3 \n" |
| 549 | "por %%xmm3,%%xmm2 \n" |
| 550 | "movdqa %%xmm1,0x10(%1) \n" |
| 551 | "movdqa %%xmm2,0x20(%1) \n" |
| 552 | "lea 0x30(%1),%1 \n" |
| 553 | "sub $0x10,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 554 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 555 | : "+r"(src), // %0 |
| 556 | "+r"(dst), // %1 |
| 557 | "+r"(pix) // %2 |
| 558 | : "m"(kShuffleMaskARGBToRGB24) // %3 |
| 559 | : "memory", "cc" |
| 560 | #if defined(__SSE2__) |
| 561 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
| 562 | #endif |
| 563 | ); |
| 564 | } |
| 565 | |
| 566 | void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 567 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 568 | "movdqa %3,%%xmm6 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 569 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 570 | "1: \n" |
| 571 | "movdqa (%0),%%xmm0 \n" |
| 572 | "movdqa 0x10(%0),%%xmm1 \n" |
| 573 | "movdqa 0x20(%0),%%xmm2 \n" |
| 574 | "movdqa 0x30(%0),%%xmm3 \n" |
| 575 | "lea 0x40(%0),%0 \n" |
| 576 | "pshufb %%xmm6,%%xmm0 \n" |
| 577 | "pshufb %%xmm6,%%xmm1 \n" |
| 578 | "pshufb %%xmm6,%%xmm2 \n" |
| 579 | "pshufb %%xmm6,%%xmm3 \n" |
| 580 | "movdqa %%xmm1,%%xmm4 \n" |
| 581 | "psrldq $0x4,%%xmm1 \n" |
| 582 | "pslldq $0xc,%%xmm4 \n" |
| 583 | "movdqa %%xmm2,%%xmm5 \n" |
| 584 | "por %%xmm4,%%xmm0 \n" |
| 585 | "pslldq $0x8,%%xmm5 \n" |
| 586 | "movdqa %%xmm0,(%1) \n" |
| 587 | "por %%xmm5,%%xmm1 \n" |
| 588 | "psrldq $0x8,%%xmm2 \n" |
| 589 | "pslldq $0x4,%%xmm3 \n" |
| 590 | "por %%xmm3,%%xmm2 \n" |
| 591 | "movdqa %%xmm1,0x10(%1) \n" |
| 592 | "movdqa %%xmm2,0x20(%1) \n" |
| 593 | "lea 0x30(%1),%1 \n" |
| 594 | "sub $0x10,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 595 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 596 | : "+r"(src), // %0 |
| 597 | "+r"(dst), // %1 |
| 598 | "+r"(pix) // %2 |
| 599 | : "m"(kShuffleMaskARGBToRAW) // %3 |
| 600 | : "memory", "cc" |
| 601 | #if defined(__SSE2__) |
| 602 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
| 603 | #endif |
| 604 | ); |
| 605 | } |
| 606 | |
fbarchard@google.com | dbcabea | 2012-10-29 21:20:25 +0000 | [diff] [blame] | 607 | void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 608 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 609 | "pcmpeqb %%xmm3,%%xmm3 \n" |
| 610 | "psrld $0x1b,%%xmm3 \n" |
| 611 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 612 | "psrld $0x1a,%%xmm4 \n" |
| 613 | "pslld $0x5,%%xmm4 \n" |
| 614 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 615 | "pslld $0xb,%%xmm5 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 616 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 617 | "1: \n" |
| 618 | "movdqa (%0),%%xmm0 \n" |
| 619 | "movdqa %%xmm0,%%xmm1 \n" |
| 620 | "movdqa %%xmm0,%%xmm2 \n" |
| 621 | "pslld $0x8,%%xmm0 \n" |
| 622 | "psrld $0x3,%%xmm1 \n" |
| 623 | "psrld $0x5,%%xmm2 \n" |
| 624 | "psrad $0x10,%%xmm0 \n" |
| 625 | "pand %%xmm3,%%xmm1 \n" |
| 626 | "pand %%xmm4,%%xmm2 \n" |
| 627 | "pand %%xmm5,%%xmm0 \n" |
| 628 | "por %%xmm2,%%xmm1 \n" |
| 629 | "por %%xmm1,%%xmm0 \n" |
| 630 | "packssdw %%xmm0,%%xmm0 \n" |
| 631 | "lea 0x10(%0),%0 \n" |
| 632 | "movq %%xmm0,(%1) \n" |
| 633 | "lea 0x8(%1),%1 \n" |
| 634 | "sub $0x4,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 635 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 636 | : "+r"(src), // %0 |
| 637 | "+r"(dst), // %1 |
| 638 | "+r"(pix) // %2 |
| 639 | : |
| 640 | : "memory", "cc" |
| 641 | #if defined(__SSE2__) |
| 642 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 643 | #endif |
| 644 | ); |
| 645 | } |
| 646 | |
| 647 | void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 648 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 649 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 650 | "psrld $0x1b,%%xmm4 \n" |
| 651 | "movdqa %%xmm4,%%xmm5 \n" |
| 652 | "pslld $0x5,%%xmm5 \n" |
| 653 | "movdqa %%xmm4,%%xmm6 \n" |
| 654 | "pslld $0xa,%%xmm6 \n" |
| 655 | "pcmpeqb %%xmm7,%%xmm7 \n" |
| 656 | "pslld $0xf,%%xmm7 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 657 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 658 | "1: \n" |
| 659 | "movdqa (%0),%%xmm0 \n" |
| 660 | "movdqa %%xmm0,%%xmm1 \n" |
| 661 | "movdqa %%xmm0,%%xmm2 \n" |
| 662 | "movdqa %%xmm0,%%xmm3 \n" |
| 663 | "psrad $0x10,%%xmm0 \n" |
| 664 | "psrld $0x3,%%xmm1 \n" |
| 665 | "psrld $0x6,%%xmm2 \n" |
| 666 | "psrld $0x9,%%xmm3 \n" |
| 667 | "pand %%xmm7,%%xmm0 \n" |
| 668 | "pand %%xmm4,%%xmm1 \n" |
| 669 | "pand %%xmm5,%%xmm2 \n" |
| 670 | "pand %%xmm6,%%xmm3 \n" |
| 671 | "por %%xmm1,%%xmm0 \n" |
| 672 | "por %%xmm3,%%xmm2 \n" |
| 673 | "por %%xmm2,%%xmm0 \n" |
| 674 | "packssdw %%xmm0,%%xmm0 \n" |
| 675 | "lea 0x10(%0),%0 \n" |
| 676 | "movq %%xmm0,(%1) \n" |
| 677 | "lea 0x8(%1),%1 \n" |
| 678 | "sub $0x4,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 679 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 680 | : "+r"(src), // %0 |
| 681 | "+r"(dst), // %1 |
| 682 | "+r"(pix) // %2 |
| 683 | : |
| 684 | : "memory", "cc" |
| 685 | #if defined(__SSE2__) |
| 686 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
| 687 | #endif |
| 688 | ); |
| 689 | } |
| 690 | |
| 691 | void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 692 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 693 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 694 | "psllw $0xc,%%xmm4 \n" |
| 695 | "movdqa %%xmm4,%%xmm3 \n" |
| 696 | "psrlw $0x8,%%xmm3 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 697 | ".p2align 4 \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 698 | "1: \n" |
| 699 | "movdqa (%0),%%xmm0 \n" |
| 700 | "movdqa %%xmm0,%%xmm1 \n" |
| 701 | "pand %%xmm3,%%xmm0 \n" |
| 702 | "pand %%xmm4,%%xmm1 \n" |
| 703 | "psrlq $0x4,%%xmm0 \n" |
| 704 | "psrlq $0x8,%%xmm1 \n" |
| 705 | "por %%xmm1,%%xmm0 \n" |
| 706 | "packuswb %%xmm0,%%xmm0 \n" |
| 707 | "lea 0x10(%0),%0 \n" |
| 708 | "movq %%xmm0,(%1) \n" |
| 709 | "lea 0x8(%1),%1 \n" |
| 710 | "sub $0x4,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 711 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 712 | : "+r"(src), // %0 |
| 713 | "+r"(dst), // %1 |
| 714 | "+r"(pix) // %2 |
| 715 | : |
| 716 | : "memory", "cc" |
| 717 | #if defined(__SSE2__) |
| 718 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" |
| 719 | #endif |
| 720 | ); |
| 721 | } |
| 722 | |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 723 | void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 724 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 725 | "movdqa %4,%%xmm5 \n" |
| 726 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 727 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 728 | "1: \n" |
| 729 | "movdqa (%0),%%xmm0 \n" |
| 730 | "movdqa 0x10(%0),%%xmm1 \n" |
| 731 | "movdqa 0x20(%0),%%xmm2 \n" |
| 732 | "movdqa 0x30(%0),%%xmm3 \n" |
| 733 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 734 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 735 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 736 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 737 | "lea 0x40(%0),%0 \n" |
| 738 | "phaddw %%xmm1,%%xmm0 \n" |
| 739 | "phaddw %%xmm3,%%xmm2 \n" |
| 740 | "psrlw $0x7,%%xmm0 \n" |
| 741 | "psrlw $0x7,%%xmm2 \n" |
| 742 | "packuswb %%xmm2,%%xmm0 \n" |
| 743 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 744 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 745 | "movdqa %%xmm0,(%1) \n" |
| 746 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 747 | "jg 1b \n" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 748 | : "+r"(src_argb), // %0 |
| 749 | "+r"(dst_y), // %1 |
| 750 | "+r"(pix) // %2 |
| 751 | : "m"(kARGBToY), // %3 |
| 752 | "m"(kAddY16) // %4 |
| 753 | : "memory", "cc" |
| 754 | #if defined(__SSE2__) |
| 755 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 756 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 757 | ); |
fbarchard@google.com | 585a126 | 2011-10-28 23:51:08 +0000 | [diff] [blame] | 758 | } |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 759 | |
| 760 | void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 761 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 762 | "movdqa %4,%%xmm5 \n" |
| 763 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 764 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 765 | "1: \n" |
| 766 | "movdqu (%0),%%xmm0 \n" |
| 767 | "movdqu 0x10(%0),%%xmm1 \n" |
| 768 | "movdqu 0x20(%0),%%xmm2 \n" |
| 769 | "movdqu 0x30(%0),%%xmm3 \n" |
| 770 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 771 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 772 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 773 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 774 | "lea 0x40(%0),%0 \n" |
| 775 | "phaddw %%xmm1,%%xmm0 \n" |
| 776 | "phaddw %%xmm3,%%xmm2 \n" |
| 777 | "psrlw $0x7,%%xmm0 \n" |
| 778 | "psrlw $0x7,%%xmm2 \n" |
| 779 | "packuswb %%xmm2,%%xmm0 \n" |
| 780 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 781 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 782 | "movdqu %%xmm0,(%1) \n" |
| 783 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 784 | "jg 1b \n" |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 785 | : "+r"(src_argb), // %0 |
| 786 | "+r"(dst_y), // %1 |
| 787 | "+r"(pix) // %2 |
| 788 | : "m"(kARGBToY), // %3 |
| 789 | "m"(kAddY16) // %4 |
| 790 | : "memory", "cc" |
| 791 | #if defined(__SSE2__) |
| 792 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 793 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 794 | ); |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 795 | } |
fbarchard@google.com | 585a126 | 2011-10-28 23:51:08 +0000 | [diff] [blame] | 796 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 797 | // TODO(fbarchard): pass xmm constants to single block of assembly. |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 798 | // fpic on GCC 4.2 for OSX runs out of GPR registers. "m" effectively takes |
| 799 | // 3 registers - ebx, ebp and eax. "m" can be passed with 3 normal registers, |
| 800 | // or 4 if stack frame is disabled. Doing 2 assembly blocks is a work around |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 801 | // and considered unsafe. |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 802 | void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb, |
| 803 | uint8* dst_u, uint8* dst_v, int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 804 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 805 | "movdqa %0,%%xmm4 \n" |
| 806 | "movdqa %1,%%xmm3 \n" |
| 807 | "movdqa %2,%%xmm5 \n" |
fbarchard@google.com | d93d448 | 2011-11-10 18:26:20 +0000 | [diff] [blame] | 808 | : |
fbarchard@google.com | f2d84dd | 2012-05-14 20:23:35 +0000 | [diff] [blame] | 809 | : "m"(kARGBToU), // %0 |
| 810 | "m"(kARGBToV), // %1 |
| 811 | "m"(kAddUV128) // %2 |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 812 | ); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 813 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 814 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 815 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 816 | "1: \n" |
| 817 | "movdqa (%0),%%xmm0 \n" |
| 818 | "movdqa 0x10(%0),%%xmm1 \n" |
| 819 | "movdqa 0x20(%0),%%xmm2 \n" |
| 820 | "movdqa 0x30(%0),%%xmm6 \n" |
| 821 | "pavgb (%0,%4,1),%%xmm0 \n" |
| 822 | "pavgb 0x10(%0,%4,1),%%xmm1 \n" |
| 823 | "pavgb 0x20(%0,%4,1),%%xmm2 \n" |
| 824 | "pavgb 0x30(%0,%4,1),%%xmm6 \n" |
| 825 | "lea 0x40(%0),%0 \n" |
| 826 | "movdqa %%xmm0,%%xmm7 \n" |
| 827 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 828 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 829 | "pavgb %%xmm7,%%xmm0 \n" |
| 830 | "movdqa %%xmm2,%%xmm7 \n" |
| 831 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 832 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 833 | "pavgb %%xmm7,%%xmm2 \n" |
| 834 | "movdqa %%xmm0,%%xmm1 \n" |
| 835 | "movdqa %%xmm2,%%xmm6 \n" |
| 836 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 837 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 838 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 839 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 840 | "phaddw %%xmm2,%%xmm0 \n" |
| 841 | "phaddw %%xmm6,%%xmm1 \n" |
| 842 | "psraw $0x8,%%xmm0 \n" |
| 843 | "psraw $0x8,%%xmm1 \n" |
| 844 | "packsswb %%xmm1,%%xmm0 \n" |
| 845 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 846 | "sub $0x10,%3 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 847 | "movlps %%xmm0,(%1) \n" |
| 848 | "movhps %%xmm0,(%1,%2,1) \n" |
| 849 | "lea 0x8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 850 | "jg 1b \n" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 851 | : "+r"(src_argb0), // %0 |
| 852 | "+r"(dst_u), // %1 |
| 853 | "+r"(dst_v), // %2 |
| 854 | "+rm"(width) // %3 |
fbarchard@google.com | d93d448 | 2011-11-10 18:26:20 +0000 | [diff] [blame] | 855 | : "r"(static_cast<intptr_t>(src_stride_argb)) |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 856 | : "memory", "cc" |
| 857 | #if defined(__SSE2__) |
fbarchard@google.com | d93d448 | 2011-11-10 18:26:20 +0000 | [diff] [blame] | 858 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 859 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 860 | ); |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 861 | } |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 862 | |
| 863 | void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb, |
| 864 | uint8* dst_u, uint8* dst_v, int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 865 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 866 | "movdqa %0,%%xmm4 \n" |
| 867 | "movdqa %1,%%xmm3 \n" |
| 868 | "movdqa %2,%%xmm5 \n" |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 869 | : |
| 870 | : "m"(kARGBToU), // %0 |
| 871 | "m"(kARGBToV), // %1 |
| 872 | "m"(kAddUV128) // %2 |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 873 | ); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 874 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 875 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 876 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 877 | "1: \n" |
| 878 | "movdqu (%0),%%xmm0 \n" |
| 879 | "movdqu 0x10(%0),%%xmm1 \n" |
| 880 | "movdqu 0x20(%0),%%xmm2 \n" |
| 881 | "movdqu 0x30(%0),%%xmm6 \n" |
| 882 | "movdqu (%0,%4,1),%%xmm7 \n" |
| 883 | "pavgb %%xmm7,%%xmm0 \n" |
| 884 | "movdqu 0x10(%0,%4,1),%%xmm7 \n" |
| 885 | "pavgb %%xmm7,%%xmm1 \n" |
| 886 | "movdqu 0x20(%0,%4,1),%%xmm7 \n" |
| 887 | "pavgb %%xmm7,%%xmm2 \n" |
| 888 | "movdqu 0x30(%0,%4,1),%%xmm7 \n" |
| 889 | "pavgb %%xmm7,%%xmm6 \n" |
| 890 | "lea 0x40(%0),%0 \n" |
| 891 | "movdqa %%xmm0,%%xmm7 \n" |
| 892 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 893 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 894 | "pavgb %%xmm7,%%xmm0 \n" |
| 895 | "movdqa %%xmm2,%%xmm7 \n" |
| 896 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 897 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 898 | "pavgb %%xmm7,%%xmm2 \n" |
| 899 | "movdqa %%xmm0,%%xmm1 \n" |
| 900 | "movdqa %%xmm2,%%xmm6 \n" |
| 901 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 902 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 903 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 904 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 905 | "phaddw %%xmm2,%%xmm0 \n" |
| 906 | "phaddw %%xmm6,%%xmm1 \n" |
| 907 | "psraw $0x8,%%xmm0 \n" |
| 908 | "psraw $0x8,%%xmm1 \n" |
| 909 | "packsswb %%xmm1,%%xmm0 \n" |
| 910 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 911 | "sub $0x10,%3 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 912 | "movlps %%xmm0,(%1) \n" |
| 913 | "movhps %%xmm0,(%1,%2,1) \n" |
| 914 | "lea 0x8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 915 | "jg 1b \n" |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 916 | : "+r"(src_argb0), // %0 |
| 917 | "+r"(dst_u), // %1 |
| 918 | "+r"(dst_v), // %2 |
| 919 | "+rm"(width) // %3 |
| 920 | : "r"(static_cast<intptr_t>(src_stride_argb)) |
| 921 | : "memory", "cc" |
| 922 | #if defined(__SSE2__) |
| 923 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 924 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 925 | ); |
fbarchard@google.com | b5b27d1 | 2012-01-28 08:44:35 +0000 | [diff] [blame] | 926 | } |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 927 | |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 928 | void ARGBToUV422Row_SSSE3(const uint8* src_argb0, |
| 929 | uint8* dst_u, uint8* dst_v, int width) { |
| 930 | asm volatile ( |
| 931 | "movdqa %0,%%xmm4 \n" |
| 932 | "movdqa %1,%%xmm3 \n" |
| 933 | "movdqa %2,%%xmm5 \n" |
| 934 | : |
| 935 | : "m"(kARGBToU), // %0 |
| 936 | "m"(kARGBToV), // %1 |
| 937 | "m"(kAddUV128) // %2 |
| 938 | ); |
| 939 | asm volatile ( |
| 940 | "sub %1,%2 \n" |
| 941 | ".p2align 4 \n" |
| 942 | "1: \n" |
| 943 | "movdqa (%0),%%xmm0 \n" |
| 944 | "movdqa 0x10(%0),%%xmm1 \n" |
| 945 | "movdqa 0x20(%0),%%xmm2 \n" |
| 946 | "movdqa 0x30(%0),%%xmm6 \n" |
| 947 | "lea 0x40(%0),%0 \n" |
| 948 | "movdqa %%xmm0,%%xmm7 \n" |
| 949 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 950 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 951 | "pavgb %%xmm7,%%xmm0 \n" |
| 952 | "movdqa %%xmm2,%%xmm7 \n" |
| 953 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 954 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 955 | "pavgb %%xmm7,%%xmm2 \n" |
| 956 | "movdqa %%xmm0,%%xmm1 \n" |
| 957 | "movdqa %%xmm2,%%xmm6 \n" |
| 958 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 959 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 960 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 961 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 962 | "phaddw %%xmm2,%%xmm0 \n" |
| 963 | "phaddw %%xmm6,%%xmm1 \n" |
| 964 | "psraw $0x8,%%xmm0 \n" |
| 965 | "psraw $0x8,%%xmm1 \n" |
| 966 | "packsswb %%xmm1,%%xmm0 \n" |
| 967 | "paddb %%xmm5,%%xmm0 \n" |
| 968 | "sub $0x10,%3 \n" |
| 969 | "movlps %%xmm0,(%1) \n" |
| 970 | "movhps %%xmm0,(%1,%2,1) \n" |
| 971 | "lea 0x8(%1),%1 \n" |
| 972 | "jg 1b \n" |
| 973 | : "+r"(src_argb0), // %0 |
| 974 | "+r"(dst_u), // %1 |
| 975 | "+r"(dst_v), // %2 |
| 976 | "+rm"(width) // %3 |
| 977 | : |
| 978 | : "memory", "cc" |
| 979 | #if defined(__SSE2__) |
| 980 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 981 | #endif |
| 982 | ); |
| 983 | } |
| 984 | |
| 985 | void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0, |
| 986 | uint8* dst_u, uint8* dst_v, int width) { |
| 987 | asm volatile ( |
| 988 | "movdqa %0,%%xmm4 \n" |
| 989 | "movdqa %1,%%xmm3 \n" |
| 990 | "movdqa %2,%%xmm5 \n" |
| 991 | : |
| 992 | : "m"(kARGBToU), // %0 |
| 993 | "m"(kARGBToV), // %1 |
| 994 | "m"(kAddUV128) // %2 |
| 995 | ); |
| 996 | asm volatile ( |
| 997 | "sub %1,%2 \n" |
| 998 | ".p2align 4 \n" |
| 999 | "1: \n" |
| 1000 | "movdqu (%0),%%xmm0 \n" |
| 1001 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1002 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1003 | "movdqu 0x30(%0),%%xmm6 \n" |
| 1004 | "lea 0x40(%0),%0 \n" |
| 1005 | "movdqa %%xmm0,%%xmm7 \n" |
| 1006 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1007 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1008 | "pavgb %%xmm7,%%xmm0 \n" |
| 1009 | "movdqa %%xmm2,%%xmm7 \n" |
| 1010 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1011 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1012 | "pavgb %%xmm7,%%xmm2 \n" |
| 1013 | "movdqa %%xmm0,%%xmm1 \n" |
| 1014 | "movdqa %%xmm2,%%xmm6 \n" |
| 1015 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1016 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1017 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1018 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1019 | "phaddw %%xmm2,%%xmm0 \n" |
| 1020 | "phaddw %%xmm6,%%xmm1 \n" |
| 1021 | "psraw $0x8,%%xmm0 \n" |
| 1022 | "psraw $0x8,%%xmm1 \n" |
| 1023 | "packsswb %%xmm1,%%xmm0 \n" |
| 1024 | "paddb %%xmm5,%%xmm0 \n" |
| 1025 | "sub $0x10,%3 \n" |
| 1026 | "movlps %%xmm0,(%1) \n" |
| 1027 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1028 | "lea 0x8(%1),%1 \n" |
| 1029 | "jg 1b \n" |
| 1030 | : "+r"(src_argb0), // %0 |
| 1031 | "+r"(dst_u), // %1 |
| 1032 | "+r"(dst_v), // %2 |
| 1033 | "+rm"(width) // %3 |
| 1034 | : |
| 1035 | : "memory", "cc" |
| 1036 | #if defined(__SSE2__) |
| 1037 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1038 | #endif |
| 1039 | ); |
| 1040 | } |
| 1041 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1042 | void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1043 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1044 | "movdqa %4,%%xmm5 \n" |
| 1045 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1046 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1047 | "1: \n" |
| 1048 | "movdqa (%0),%%xmm0 \n" |
| 1049 | "movdqa 0x10(%0),%%xmm1 \n" |
| 1050 | "movdqa 0x20(%0),%%xmm2 \n" |
| 1051 | "movdqa 0x30(%0),%%xmm3 \n" |
| 1052 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1053 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 1054 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1055 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 1056 | "lea 0x40(%0),%0 \n" |
| 1057 | "phaddw %%xmm1,%%xmm0 \n" |
| 1058 | "phaddw %%xmm3,%%xmm2 \n" |
| 1059 | "psrlw $0x7,%%xmm0 \n" |
| 1060 | "psrlw $0x7,%%xmm2 \n" |
| 1061 | "packuswb %%xmm2,%%xmm0 \n" |
| 1062 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1063 | "sub $0x10,%2 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1064 | "movdqa %%xmm0,(%1) \n" |
| 1065 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1066 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1067 | : "+r"(src_bgra), // %0 |
| 1068 | "+r"(dst_y), // %1 |
| 1069 | "+r"(pix) // %2 |
| 1070 | : "m"(kBGRAToY), // %3 |
| 1071 | "m"(kAddY16) // %4 |
| 1072 | : "memory", "cc" |
| 1073 | #if defined(__SSE2__) |
| 1074 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 1075 | #endif |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1076 | ); |
| 1077 | } |
| 1078 | |
| 1079 | void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1080 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1081 | "movdqa %4,%%xmm5 \n" |
| 1082 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1083 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1084 | "1: \n" |
| 1085 | "movdqu (%0),%%xmm0 \n" |
| 1086 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1087 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1088 | "movdqu 0x30(%0),%%xmm3 \n" |
| 1089 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1090 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 1091 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1092 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 1093 | "lea 0x40(%0),%0 \n" |
| 1094 | "phaddw %%xmm1,%%xmm0 \n" |
| 1095 | "phaddw %%xmm3,%%xmm2 \n" |
| 1096 | "psrlw $0x7,%%xmm0 \n" |
| 1097 | "psrlw $0x7,%%xmm2 \n" |
| 1098 | "packuswb %%xmm2,%%xmm0 \n" |
| 1099 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1100 | "sub $0x10,%2 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1101 | "movdqu %%xmm0,(%1) \n" |
| 1102 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1103 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1104 | : "+r"(src_bgra), // %0 |
| 1105 | "+r"(dst_y), // %1 |
| 1106 | "+r"(pix) // %2 |
| 1107 | : "m"(kBGRAToY), // %3 |
| 1108 | "m"(kAddY16) // %4 |
| 1109 | : "memory", "cc" |
| 1110 | #if defined(__SSE2__) |
| 1111 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1112 | #endif |
| 1113 | ); |
| 1114 | } |
| 1115 | |
| 1116 | void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra, |
| 1117 | uint8* dst_u, uint8* dst_v, int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1118 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1119 | "movdqa %0,%%xmm4 \n" |
| 1120 | "movdqa %1,%%xmm3 \n" |
| 1121 | "movdqa %2,%%xmm5 \n" |
| 1122 | : |
| 1123 | : "m"(kBGRAToU), // %0 |
| 1124 | "m"(kBGRAToV), // %1 |
| 1125 | "m"(kAddUV128) // %2 |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1126 | ); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1127 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1128 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1129 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1130 | "1: \n" |
| 1131 | "movdqa (%0),%%xmm0 \n" |
| 1132 | "movdqa 0x10(%0),%%xmm1 \n" |
| 1133 | "movdqa 0x20(%0),%%xmm2 \n" |
| 1134 | "movdqa 0x30(%0),%%xmm6 \n" |
| 1135 | "pavgb (%0,%4,1),%%xmm0 \n" |
| 1136 | "pavgb 0x10(%0,%4,1),%%xmm1 \n" |
| 1137 | "pavgb 0x20(%0,%4,1),%%xmm2 \n" |
| 1138 | "pavgb 0x30(%0,%4,1),%%xmm6 \n" |
| 1139 | "lea 0x40(%0),%0 \n" |
| 1140 | "movdqa %%xmm0,%%xmm7 \n" |
| 1141 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1142 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1143 | "pavgb %%xmm7,%%xmm0 \n" |
| 1144 | "movdqa %%xmm2,%%xmm7 \n" |
| 1145 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1146 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1147 | "pavgb %%xmm7,%%xmm2 \n" |
| 1148 | "movdqa %%xmm0,%%xmm1 \n" |
| 1149 | "movdqa %%xmm2,%%xmm6 \n" |
| 1150 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1151 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1152 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1153 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1154 | "phaddw %%xmm2,%%xmm0 \n" |
| 1155 | "phaddw %%xmm6,%%xmm1 \n" |
| 1156 | "psraw $0x8,%%xmm0 \n" |
| 1157 | "psraw $0x8,%%xmm1 \n" |
| 1158 | "packsswb %%xmm1,%%xmm0 \n" |
| 1159 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1160 | "sub $0x10,%3 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1161 | "movlps %%xmm0,(%1) \n" |
| 1162 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1163 | "lea 0x8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1164 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1165 | : "+r"(src_bgra0), // %0 |
| 1166 | "+r"(dst_u), // %1 |
| 1167 | "+r"(dst_v), // %2 |
| 1168 | "+rm"(width) // %3 |
| 1169 | : "r"(static_cast<intptr_t>(src_stride_bgra)) |
| 1170 | : "memory", "cc" |
| 1171 | #if defined(__SSE2__) |
| 1172 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1173 | #endif |
| 1174 | ); |
| 1175 | } |
| 1176 | |
| 1177 | void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra0, int src_stride_bgra, |
| 1178 | uint8* dst_u, uint8* dst_v, int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1179 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1180 | "movdqa %0,%%xmm4 \n" |
| 1181 | "movdqa %1,%%xmm3 \n" |
| 1182 | "movdqa %2,%%xmm5 \n" |
| 1183 | : |
| 1184 | : "m"(kBGRAToU), // %0 |
| 1185 | "m"(kBGRAToV), // %1 |
| 1186 | "m"(kAddUV128) // %2 |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1187 | ); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1188 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1189 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1190 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1191 | "1: \n" |
| 1192 | "movdqu (%0),%%xmm0 \n" |
| 1193 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1194 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1195 | "movdqu 0x30(%0),%%xmm6 \n" |
| 1196 | "movdqu (%0,%4,1),%%xmm7 \n" |
| 1197 | "pavgb %%xmm7,%%xmm0 \n" |
| 1198 | "movdqu 0x10(%0,%4,1),%%xmm7 \n" |
| 1199 | "pavgb %%xmm7,%%xmm1 \n" |
| 1200 | "movdqu 0x20(%0,%4,1),%%xmm7 \n" |
| 1201 | "pavgb %%xmm7,%%xmm2 \n" |
| 1202 | "movdqu 0x30(%0,%4,1),%%xmm7 \n" |
| 1203 | "pavgb %%xmm7,%%xmm6 \n" |
| 1204 | "lea 0x40(%0),%0 \n" |
| 1205 | "movdqa %%xmm0,%%xmm7 \n" |
| 1206 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1207 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1208 | "pavgb %%xmm7,%%xmm0 \n" |
| 1209 | "movdqa %%xmm2,%%xmm7 \n" |
| 1210 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1211 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1212 | "pavgb %%xmm7,%%xmm2 \n" |
| 1213 | "movdqa %%xmm0,%%xmm1 \n" |
| 1214 | "movdqa %%xmm2,%%xmm6 \n" |
| 1215 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1216 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1217 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1218 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1219 | "phaddw %%xmm2,%%xmm0 \n" |
| 1220 | "phaddw %%xmm6,%%xmm1 \n" |
| 1221 | "psraw $0x8,%%xmm0 \n" |
| 1222 | "psraw $0x8,%%xmm1 \n" |
| 1223 | "packsswb %%xmm1,%%xmm0 \n" |
| 1224 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1225 | "sub $0x10,%3 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1226 | "movlps %%xmm0,(%1) \n" |
| 1227 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1228 | "lea 0x8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1229 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1230 | : "+r"(src_bgra0), // %0 |
| 1231 | "+r"(dst_u), // %1 |
| 1232 | "+r"(dst_v), // %2 |
| 1233 | "+rm"(width) // %3 |
| 1234 | : "r"(static_cast<intptr_t>(src_stride_bgra)) |
| 1235 | : "memory", "cc" |
| 1236 | #if defined(__SSE2__) |
| 1237 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1238 | #endif |
| 1239 | ); |
| 1240 | } |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1241 | |
| 1242 | void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1243 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1244 | "movdqa %4,%%xmm5 \n" |
| 1245 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1246 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1247 | "1: \n" |
| 1248 | "movdqa (%0),%%xmm0 \n" |
| 1249 | "movdqa 0x10(%0),%%xmm1 \n" |
| 1250 | "movdqa 0x20(%0),%%xmm2 \n" |
| 1251 | "movdqa 0x30(%0),%%xmm3 \n" |
| 1252 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1253 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 1254 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1255 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 1256 | "lea 0x40(%0),%0 \n" |
| 1257 | "phaddw %%xmm1,%%xmm0 \n" |
| 1258 | "phaddw %%xmm3,%%xmm2 \n" |
| 1259 | "psrlw $0x7,%%xmm0 \n" |
| 1260 | "psrlw $0x7,%%xmm2 \n" |
| 1261 | "packuswb %%xmm2,%%xmm0 \n" |
| 1262 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1263 | "sub $0x10,%2 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1264 | "movdqa %%xmm0,(%1) \n" |
| 1265 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1266 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1267 | : "+r"(src_abgr), // %0 |
| 1268 | "+r"(dst_y), // %1 |
| 1269 | "+r"(pix) // %2 |
| 1270 | : "m"(kABGRToY), // %3 |
| 1271 | "m"(kAddY16) // %4 |
| 1272 | : "memory", "cc" |
| 1273 | #if defined(__SSE2__) |
| 1274 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1275 | #endif |
| 1276 | ); |
| 1277 | } |
| 1278 | |
| 1279 | void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1280 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1281 | "movdqa %4,%%xmm5 \n" |
| 1282 | "movdqa %3,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1283 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1284 | "1: \n" |
| 1285 | "movdqu (%0),%%xmm0 \n" |
| 1286 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1287 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1288 | "movdqu 0x30(%0),%%xmm3 \n" |
| 1289 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1290 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 1291 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1292 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 1293 | "lea 0x40(%0),%0 \n" |
| 1294 | "phaddw %%xmm1,%%xmm0 \n" |
| 1295 | "phaddw %%xmm3,%%xmm2 \n" |
| 1296 | "psrlw $0x7,%%xmm0 \n" |
| 1297 | "psrlw $0x7,%%xmm2 \n" |
| 1298 | "packuswb %%xmm2,%%xmm0 \n" |
| 1299 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1300 | "sub $0x10,%2 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1301 | "movdqu %%xmm0,(%1) \n" |
| 1302 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1303 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1304 | : "+r"(src_abgr), // %0 |
| 1305 | "+r"(dst_y), // %1 |
| 1306 | "+r"(pix) // %2 |
| 1307 | : "m"(kABGRToY), // %3 |
| 1308 | "m"(kAddY16) // %4 |
| 1309 | : "memory", "cc" |
| 1310 | #if defined(__SSE2__) |
| 1311 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1312 | #endif |
| 1313 | ); |
| 1314 | } |
| 1315 | |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1316 | void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) { |
| 1317 | asm volatile ( |
| 1318 | "movdqa %4,%%xmm5 \n" |
| 1319 | "movdqa %3,%%xmm4 \n" |
| 1320 | ".p2align 4 \n" |
| 1321 | "1: \n" |
| 1322 | "movdqa (%0),%%xmm0 \n" |
| 1323 | "movdqa 0x10(%0),%%xmm1 \n" |
| 1324 | "movdqa 0x20(%0),%%xmm2 \n" |
| 1325 | "movdqa 0x30(%0),%%xmm3 \n" |
| 1326 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1327 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 1328 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1329 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 1330 | "lea 0x40(%0),%0 \n" |
| 1331 | "phaddw %%xmm1,%%xmm0 \n" |
| 1332 | "phaddw %%xmm3,%%xmm2 \n" |
| 1333 | "psrlw $0x7,%%xmm0 \n" |
| 1334 | "psrlw $0x7,%%xmm2 \n" |
| 1335 | "packuswb %%xmm2,%%xmm0 \n" |
| 1336 | "paddb %%xmm5,%%xmm0 \n" |
| 1337 | "sub $0x10,%2 \n" |
| 1338 | "movdqa %%xmm0,(%1) \n" |
| 1339 | "lea 0x10(%1),%1 \n" |
| 1340 | "jg 1b \n" |
| 1341 | : "+r"(src_rgba), // %0 |
| 1342 | "+r"(dst_y), // %1 |
| 1343 | "+r"(pix) // %2 |
| 1344 | : "m"(kRGBAToY), // %3 |
| 1345 | "m"(kAddY16) // %4 |
| 1346 | : "memory", "cc" |
| 1347 | #if defined(__SSE2__) |
| 1348 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1349 | #endif |
| 1350 | ); |
| 1351 | } |
| 1352 | |
| 1353 | void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) { |
| 1354 | asm volatile ( |
| 1355 | "movdqa %4,%%xmm5 \n" |
| 1356 | "movdqa %3,%%xmm4 \n" |
| 1357 | ".p2align 4 \n" |
| 1358 | "1: \n" |
| 1359 | "movdqu (%0),%%xmm0 \n" |
| 1360 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1361 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1362 | "movdqu 0x30(%0),%%xmm3 \n" |
| 1363 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1364 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 1365 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1366 | "pmaddubsw %%xmm4,%%xmm3 \n" |
| 1367 | "lea 0x40(%0),%0 \n" |
| 1368 | "phaddw %%xmm1,%%xmm0 \n" |
| 1369 | "phaddw %%xmm3,%%xmm2 \n" |
| 1370 | "psrlw $0x7,%%xmm0 \n" |
| 1371 | "psrlw $0x7,%%xmm2 \n" |
| 1372 | "packuswb %%xmm2,%%xmm0 \n" |
| 1373 | "paddb %%xmm5,%%xmm0 \n" |
| 1374 | "sub $0x10,%2 \n" |
| 1375 | "movdqu %%xmm0,(%1) \n" |
| 1376 | "lea 0x10(%1),%1 \n" |
| 1377 | "jg 1b \n" |
| 1378 | : "+r"(src_rgba), // %0 |
| 1379 | "+r"(dst_y), // %1 |
| 1380 | "+r"(pix) // %2 |
| 1381 | : "m"(kRGBAToY), // %3 |
| 1382 | "m"(kAddY16) // %4 |
| 1383 | : "memory", "cc" |
| 1384 | #if defined(__SSE2__) |
| 1385 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1386 | #endif |
| 1387 | ); |
| 1388 | } |
| 1389 | |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1390 | void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr, |
| 1391 | uint8* dst_u, uint8* dst_v, int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1392 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1393 | "movdqa %0,%%xmm4 \n" |
| 1394 | "movdqa %1,%%xmm3 \n" |
| 1395 | "movdqa %2,%%xmm5 \n" |
| 1396 | : |
| 1397 | : "m"(kABGRToU), // %0 |
| 1398 | "m"(kABGRToV), // %1 |
| 1399 | "m"(kAddUV128) // %2 |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1400 | ); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1401 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1402 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1403 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1404 | "1: \n" |
| 1405 | "movdqa (%0),%%xmm0 \n" |
| 1406 | "movdqa 0x10(%0),%%xmm1 \n" |
| 1407 | "movdqa 0x20(%0),%%xmm2 \n" |
| 1408 | "movdqa 0x30(%0),%%xmm6 \n" |
| 1409 | "pavgb (%0,%4,1),%%xmm0 \n" |
| 1410 | "pavgb 0x10(%0,%4,1),%%xmm1 \n" |
| 1411 | "pavgb 0x20(%0,%4,1),%%xmm2 \n" |
| 1412 | "pavgb 0x30(%0,%4,1),%%xmm6 \n" |
| 1413 | "lea 0x40(%0),%0 \n" |
| 1414 | "movdqa %%xmm0,%%xmm7 \n" |
| 1415 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1416 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1417 | "pavgb %%xmm7,%%xmm0 \n" |
| 1418 | "movdqa %%xmm2,%%xmm7 \n" |
| 1419 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1420 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1421 | "pavgb %%xmm7,%%xmm2 \n" |
| 1422 | "movdqa %%xmm0,%%xmm1 \n" |
| 1423 | "movdqa %%xmm2,%%xmm6 \n" |
| 1424 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1425 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1426 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1427 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1428 | "phaddw %%xmm2,%%xmm0 \n" |
| 1429 | "phaddw %%xmm6,%%xmm1 \n" |
| 1430 | "psraw $0x8,%%xmm0 \n" |
| 1431 | "psraw $0x8,%%xmm1 \n" |
| 1432 | "packsswb %%xmm1,%%xmm0 \n" |
| 1433 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1434 | "sub $0x10,%3 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1435 | "movlps %%xmm0,(%1) \n" |
| 1436 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1437 | "lea 0x8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1438 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1439 | : "+r"(src_abgr0), // %0 |
| 1440 | "+r"(dst_u), // %1 |
| 1441 | "+r"(dst_v), // %2 |
| 1442 | "+rm"(width) // %3 |
| 1443 | : "r"(static_cast<intptr_t>(src_stride_abgr)) |
| 1444 | : "memory", "cc" |
| 1445 | #if defined(__SSE2__) |
| 1446 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1447 | #endif |
| 1448 | ); |
| 1449 | } |
| 1450 | |
| 1451 | void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr0, int src_stride_abgr, |
| 1452 | uint8* dst_u, uint8* dst_v, int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1453 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1454 | "movdqa %0,%%xmm4 \n" |
| 1455 | "movdqa %1,%%xmm3 \n" |
| 1456 | "movdqa %2,%%xmm5 \n" |
| 1457 | : |
| 1458 | : "m"(kABGRToU), // %0 |
| 1459 | "m"(kABGRToV), // %1 |
| 1460 | "m"(kAddUV128) // %2 |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1461 | ); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1462 | asm volatile ( |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1463 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1464 | ".p2align 4 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1465 | "1: \n" |
| 1466 | "movdqu (%0),%%xmm0 \n" |
| 1467 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1468 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1469 | "movdqu 0x30(%0),%%xmm6 \n" |
| 1470 | "movdqu (%0,%4,1),%%xmm7 \n" |
| 1471 | "pavgb %%xmm7,%%xmm0 \n" |
| 1472 | "movdqu 0x10(%0,%4,1),%%xmm7 \n" |
| 1473 | "pavgb %%xmm7,%%xmm1 \n" |
| 1474 | "movdqu 0x20(%0,%4,1),%%xmm7 \n" |
| 1475 | "pavgb %%xmm7,%%xmm2 \n" |
| 1476 | "movdqu 0x30(%0,%4,1),%%xmm7 \n" |
| 1477 | "pavgb %%xmm7,%%xmm6 \n" |
| 1478 | "lea 0x40(%0),%0 \n" |
| 1479 | "movdqa %%xmm0,%%xmm7 \n" |
| 1480 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1481 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1482 | "pavgb %%xmm7,%%xmm0 \n" |
| 1483 | "movdqa %%xmm2,%%xmm7 \n" |
| 1484 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1485 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1486 | "pavgb %%xmm7,%%xmm2 \n" |
| 1487 | "movdqa %%xmm0,%%xmm1 \n" |
| 1488 | "movdqa %%xmm2,%%xmm6 \n" |
| 1489 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1490 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1491 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1492 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1493 | "phaddw %%xmm2,%%xmm0 \n" |
| 1494 | "phaddw %%xmm6,%%xmm1 \n" |
| 1495 | "psraw $0x8,%%xmm0 \n" |
| 1496 | "psraw $0x8,%%xmm1 \n" |
| 1497 | "packsswb %%xmm1,%%xmm0 \n" |
| 1498 | "paddb %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1499 | "sub $0x10,%3 \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1500 | "movlps %%xmm0,(%1) \n" |
| 1501 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1502 | "lea 0x8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1503 | "jg 1b \n" |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1504 | : "+r"(src_abgr0), // %0 |
| 1505 | "+r"(dst_u), // %1 |
| 1506 | "+r"(dst_v), // %2 |
| 1507 | "+rm"(width) // %3 |
| 1508 | : "r"(static_cast<intptr_t>(src_stride_abgr)) |
| 1509 | : "memory", "cc" |
| 1510 | #if defined(__SSE2__) |
| 1511 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1512 | #endif |
| 1513 | ); |
| 1514 | } |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1515 | |
| 1516 | void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba, |
| 1517 | uint8* dst_u, uint8* dst_v, int width) { |
| 1518 | asm volatile ( |
| 1519 | "movdqa %0,%%xmm4 \n" |
| 1520 | "movdqa %1,%%xmm3 \n" |
| 1521 | "movdqa %2,%%xmm5 \n" |
| 1522 | : |
| 1523 | : "m"(kRGBAToU), // %0 |
| 1524 | "m"(kRGBAToV), // %1 |
| 1525 | "m"(kAddUV128) // %2 |
| 1526 | ); |
| 1527 | asm volatile ( |
| 1528 | "sub %1,%2 \n" |
| 1529 | ".p2align 4 \n" |
| 1530 | "1: \n" |
| 1531 | "movdqa (%0),%%xmm0 \n" |
| 1532 | "movdqa 0x10(%0),%%xmm1 \n" |
| 1533 | "movdqa 0x20(%0),%%xmm2 \n" |
| 1534 | "movdqa 0x30(%0),%%xmm6 \n" |
| 1535 | "pavgb (%0,%4,1),%%xmm0 \n" |
| 1536 | "pavgb 0x10(%0,%4,1),%%xmm1 \n" |
| 1537 | "pavgb 0x20(%0,%4,1),%%xmm2 \n" |
| 1538 | "pavgb 0x30(%0,%4,1),%%xmm6 \n" |
| 1539 | "lea 0x40(%0),%0 \n" |
| 1540 | "movdqa %%xmm0,%%xmm7 \n" |
| 1541 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1542 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1543 | "pavgb %%xmm7,%%xmm0 \n" |
| 1544 | "movdqa %%xmm2,%%xmm7 \n" |
| 1545 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1546 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1547 | "pavgb %%xmm7,%%xmm2 \n" |
| 1548 | "movdqa %%xmm0,%%xmm1 \n" |
| 1549 | "movdqa %%xmm2,%%xmm6 \n" |
| 1550 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1551 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1552 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1553 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1554 | "phaddw %%xmm2,%%xmm0 \n" |
| 1555 | "phaddw %%xmm6,%%xmm1 \n" |
| 1556 | "psraw $0x8,%%xmm0 \n" |
| 1557 | "psraw $0x8,%%xmm1 \n" |
| 1558 | "packsswb %%xmm1,%%xmm0 \n" |
| 1559 | "paddb %%xmm5,%%xmm0 \n" |
| 1560 | "sub $0x10,%3 \n" |
| 1561 | "movlps %%xmm0,(%1) \n" |
| 1562 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1563 | "lea 0x8(%1),%1 \n" |
| 1564 | "jg 1b \n" |
| 1565 | : "+r"(src_rgba0), // %0 |
| 1566 | "+r"(dst_u), // %1 |
| 1567 | "+r"(dst_v), // %2 |
| 1568 | "+rm"(width) // %3 |
| 1569 | : "r"(static_cast<intptr_t>(src_stride_rgba)) |
| 1570 | : "memory", "cc" |
| 1571 | #if defined(__SSE2__) |
| 1572 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1573 | #endif |
| 1574 | ); |
| 1575 | } |
| 1576 | |
| 1577 | void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba0, int src_stride_rgba, |
| 1578 | uint8* dst_u, uint8* dst_v, int width) { |
| 1579 | asm volatile ( |
| 1580 | "movdqa %0,%%xmm4 \n" |
| 1581 | "movdqa %1,%%xmm3 \n" |
| 1582 | "movdqa %2,%%xmm5 \n" |
| 1583 | : |
| 1584 | : "m"(kRGBAToU), // %0 |
| 1585 | "m"(kRGBAToV), // %1 |
| 1586 | "m"(kAddUV128) // %2 |
| 1587 | ); |
| 1588 | asm volatile ( |
| 1589 | "sub %1,%2 \n" |
| 1590 | ".p2align 4 \n" |
| 1591 | "1: \n" |
| 1592 | "movdqu (%0),%%xmm0 \n" |
| 1593 | "movdqu 0x10(%0),%%xmm1 \n" |
| 1594 | "movdqu 0x20(%0),%%xmm2 \n" |
| 1595 | "movdqu 0x30(%0),%%xmm6 \n" |
| 1596 | "movdqu (%0,%4,1),%%xmm7 \n" |
| 1597 | "pavgb %%xmm7,%%xmm0 \n" |
| 1598 | "movdqu 0x10(%0,%4,1),%%xmm7 \n" |
| 1599 | "pavgb %%xmm7,%%xmm1 \n" |
| 1600 | "movdqu 0x20(%0,%4,1),%%xmm7 \n" |
| 1601 | "pavgb %%xmm7,%%xmm2 \n" |
| 1602 | "movdqu 0x30(%0,%4,1),%%xmm7 \n" |
| 1603 | "pavgb %%xmm7,%%xmm6 \n" |
| 1604 | "lea 0x40(%0),%0 \n" |
| 1605 | "movdqa %%xmm0,%%xmm7 \n" |
| 1606 | "shufps $0x88,%%xmm1,%%xmm0 \n" |
| 1607 | "shufps $0xdd,%%xmm1,%%xmm7 \n" |
| 1608 | "pavgb %%xmm7,%%xmm0 \n" |
| 1609 | "movdqa %%xmm2,%%xmm7 \n" |
| 1610 | "shufps $0x88,%%xmm6,%%xmm2 \n" |
| 1611 | "shufps $0xdd,%%xmm6,%%xmm7 \n" |
| 1612 | "pavgb %%xmm7,%%xmm2 \n" |
| 1613 | "movdqa %%xmm0,%%xmm1 \n" |
| 1614 | "movdqa %%xmm2,%%xmm6 \n" |
| 1615 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 1616 | "pmaddubsw %%xmm4,%%xmm2 \n" |
| 1617 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 1618 | "pmaddubsw %%xmm3,%%xmm6 \n" |
| 1619 | "phaddw %%xmm2,%%xmm0 \n" |
| 1620 | "phaddw %%xmm6,%%xmm1 \n" |
| 1621 | "psraw $0x8,%%xmm0 \n" |
| 1622 | "psraw $0x8,%%xmm1 \n" |
| 1623 | "packsswb %%xmm1,%%xmm0 \n" |
| 1624 | "paddb %%xmm5,%%xmm0 \n" |
| 1625 | "sub $0x10,%3 \n" |
| 1626 | "movlps %%xmm0,(%1) \n" |
| 1627 | "movhps %%xmm0,(%1,%2,1) \n" |
| 1628 | "lea 0x8(%1),%1 \n" |
| 1629 | "jg 1b \n" |
| 1630 | : "+r"(src_rgba0), // %0 |
| 1631 | "+r"(dst_u), // %1 |
| 1632 | "+r"(dst_v), // %2 |
| 1633 | "+rm"(width) // %3 |
| 1634 | : "r"(static_cast<intptr_t>(src_stride_rgba)) |
| 1635 | : "memory", "cc" |
| 1636 | #if defined(__SSE2__) |
| 1637 | , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" |
| 1638 | #endif |
| 1639 | ); |
| 1640 | } |
fbarchard@google.com | 714050a | 2012-02-17 22:59:56 +0000 | [diff] [blame] | 1641 | #endif // HAS_ARGBTOYROW_SSSE3 |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 1642 | |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1643 | #ifdef HAS_I422TOARGBROW_SSSE3 |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 1644 | #define UB 127 /* min(63,static_cast<int8>(2.018 * 64)) */ |
| 1645 | #define UG -25 /* static_cast<int8>(-0.391 * 64 - 0.5) */ |
| 1646 | #define UR 0 |
| 1647 | |
| 1648 | #define VB 0 |
| 1649 | #define VG -52 /* static_cast<int8>(-0.813 * 64 - 0.5) */ |
| 1650 | #define VR 102 /* static_cast<int8>(1.596 * 64 + 0.5) */ |
| 1651 | |
| 1652 | // Bias |
| 1653 | #define BB UB * 128 + VB * 128 |
| 1654 | #define BG UG * 128 + VG * 128 |
| 1655 | #define BR UR * 128 + VR * 128 |
| 1656 | |
fbarchard@google.com | 1c2d8be | 2011-11-17 21:57:54 +0000 | [diff] [blame] | 1657 | #define YG 74 /* static_cast<int8>(1.164 * 64 + 0.5) */ |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 1658 | |
fbarchard@google.com | 1c2d8be | 2011-11-17 21:57:54 +0000 | [diff] [blame] | 1659 | struct { |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1660 | vec8 kUVToB; // 0 |
| 1661 | vec8 kUVToG; // 16 |
| 1662 | vec8 kUVToR; // 32 |
| 1663 | vec16 kUVBiasB; // 48 |
| 1664 | vec16 kUVBiasG; // 64 |
| 1665 | vec16 kUVBiasR; // 80 |
| 1666 | vec16 kYSub16; // 96 |
| 1667 | vec16 kYToRgb; // 112 |
| 1668 | vec8 kVUToB; // 128 |
| 1669 | vec8 kVUToG; // 144 |
| 1670 | vec8 kVUToR; // 160 |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 1671 | } CONST SIMD_ALIGNED(kYuvConstants) = { |
fbarchard@google.com | 1c2d8be | 2011-11-17 21:57:54 +0000 | [diff] [blame] | 1672 | { UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB }, |
| 1673 | { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG }, |
| 1674 | { UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR }, |
| 1675 | { BB, BB, BB, BB, BB, BB, BB, BB }, |
| 1676 | { BG, BG, BG, BG, BG, BG, BG, BG }, |
| 1677 | { BR, BR, BR, BR, BR, BR, BR, BR }, |
| 1678 | { 16, 16, 16, 16, 16, 16, 16, 16 }, |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1679 | { YG, YG, YG, YG, YG, YG, YG, YG }, |
| 1680 | { VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB }, |
| 1681 | { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG }, |
| 1682 | { VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR } |
fbarchard@google.com | 1c2d8be | 2011-11-17 21:57:54 +0000 | [diff] [blame] | 1683 | }; |
| 1684 | |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1685 | |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1686 | // Read 8 UV from 411 |
| 1687 | #define READYUV444 \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1688 | "movq (%[u_buf]),%%xmm0 \n" \ |
| 1689 | "movq (%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
| 1690 | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1691 | "punpcklbw %%xmm1,%%xmm0 \n" \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1692 | |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1693 | // Read 4 UV from 422, upsample to 8 UV |
| 1694 | #define READYUV422 \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1695 | "movd (%[u_buf]),%%xmm0 \n" \ |
| 1696 | "movd (%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
| 1697 | "lea 0x4(%[u_buf]),%[u_buf] \n" \ |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1698 | "punpcklbw %%xmm1,%%xmm0 \n" \ |
| 1699 | "punpcklwd %%xmm0,%%xmm0 \n" \ |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 1700 | |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1701 | // Read 2 UV from 411, upsample to 8 UV |
| 1702 | #define READYUV411 \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1703 | "movd (%[u_buf]),%%xmm0 \n" \ |
| 1704 | "movd (%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
| 1705 | "lea 0x2(%[u_buf]),%[u_buf] \n" \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1706 | "punpcklbw %%xmm1,%%xmm0 \n" \ |
| 1707 | "punpcklwd %%xmm0,%%xmm0 \n" \ |
| 1708 | "punpckldq %%xmm0,%%xmm0 \n" \ |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1709 | |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1710 | // Read 4 UV from NV12, upsample to 8 UV |
| 1711 | #define READNV12 \ |
| 1712 | "movq (%[uv_buf]),%%xmm0 \n" \ |
| 1713 | "lea 0x8(%[uv_buf]),%[uv_buf] \n" \ |
fbarchard@google.com | 6fd84a8 | 2012-09-19 07:35:45 +0000 | [diff] [blame] | 1714 | "punpcklwd %%xmm0,%%xmm0 \n" \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1715 | |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1716 | // Convert 8 pixels: 8 UV and 8 Y |
| 1717 | #define YUVTORGB \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1718 | "movdqa %%xmm0,%%xmm1 \n" \ |
| 1719 | "movdqa %%xmm0,%%xmm2 \n" \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1720 | "pmaddubsw (%[kYuvConstants]),%%xmm0 \n" \ |
| 1721 | "pmaddubsw 16(%[kYuvConstants]),%%xmm1 \n" \ |
| 1722 | "pmaddubsw 32(%[kYuvConstants]),%%xmm2 \n" \ |
| 1723 | "psubw 48(%[kYuvConstants]),%%xmm0 \n" \ |
| 1724 | "psubw 64(%[kYuvConstants]),%%xmm1 \n" \ |
| 1725 | "psubw 80(%[kYuvConstants]),%%xmm2 \n" \ |
| 1726 | "movq (%[y_buf]),%%xmm3 \n" \ |
| 1727 | "lea 0x8(%[y_buf]),%[y_buf] \n" \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1728 | "punpcklbw %%xmm4,%%xmm3 \n" \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1729 | "psubsw 96(%[kYuvConstants]),%%xmm3 \n" \ |
| 1730 | "pmullw 112(%[kYuvConstants]),%%xmm3 \n" \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1731 | "paddsw %%xmm3,%%xmm0 \n" \ |
| 1732 | "paddsw %%xmm3,%%xmm1 \n" \ |
| 1733 | "paddsw %%xmm3,%%xmm2 \n" \ |
| 1734 | "psraw $0x6,%%xmm0 \n" \ |
| 1735 | "psraw $0x6,%%xmm1 \n" \ |
| 1736 | "psraw $0x6,%%xmm2 \n" \ |
| 1737 | "packuswb %%xmm0,%%xmm0 \n" \ |
| 1738 | "packuswb %%xmm1,%%xmm1 \n" \ |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1739 | "packuswb %%xmm2,%%xmm2 \n" \ |
| 1740 | |
| 1741 | // Convert 8 pixels: 8 VU and 8 Y |
| 1742 | #define YVUTORGB \ |
| 1743 | "movdqa %%xmm0,%%xmm1 \n" \ |
| 1744 | "movdqa %%xmm0,%%xmm2 \n" \ |
| 1745 | "pmaddubsw 128(%[kYuvConstants]),%%xmm0 \n" \ |
| 1746 | "pmaddubsw 144(%[kYuvConstants]),%%xmm1 \n" \ |
| 1747 | "pmaddubsw 160(%[kYuvConstants]),%%xmm2 \n" \ |
| 1748 | "psubw 48(%[kYuvConstants]),%%xmm0 \n" \ |
| 1749 | "psubw 64(%[kYuvConstants]),%%xmm1 \n" \ |
| 1750 | "psubw 80(%[kYuvConstants]),%%xmm2 \n" \ |
| 1751 | "movq (%[y_buf]),%%xmm3 \n" \ |
| 1752 | "lea 0x8(%[y_buf]),%[y_buf] \n" \ |
| 1753 | "punpcklbw %%xmm4,%%xmm3 \n" \ |
| 1754 | "psubsw 96(%[kYuvConstants]),%%xmm3 \n" \ |
| 1755 | "pmullw 112(%[kYuvConstants]),%%xmm3 \n" \ |
| 1756 | "paddsw %%xmm3,%%xmm0 \n" \ |
| 1757 | "paddsw %%xmm3,%%xmm1 \n" \ |
| 1758 | "paddsw %%xmm3,%%xmm2 \n" \ |
| 1759 | "psraw $0x6,%%xmm0 \n" \ |
| 1760 | "psraw $0x6,%%xmm1 \n" \ |
| 1761 | "psraw $0x6,%%xmm2 \n" \ |
| 1762 | "packuswb %%xmm0,%%xmm0 \n" \ |
| 1763 | "packuswb %%xmm1,%%xmm1 \n" \ |
| 1764 | "packuswb %%xmm2,%%xmm2 \n" \ |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1765 | |
| 1766 | void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf, |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 1767 | const uint8* u_buf, |
| 1768 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 1769 | uint8* dst_argb, |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 1770 | int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1771 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1772 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1773 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 1774 | "pxor %%xmm4,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1775 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1776 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1777 | READYUV444 |
| 1778 | YUVTORGB |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1779 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 1780 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 1781 | "movdqa %%xmm0,%%xmm1 \n" |
| 1782 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 1783 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1784 | "movdqa %%xmm0,(%[dst_argb]) \n" |
| 1785 | "movdqa %%xmm1,0x10(%[dst_argb]) \n" |
| 1786 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1787 | "sub $0x8,%[width] \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 1788 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1789 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 1790 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 1791 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1792 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1793 | [width]"+rm"(width) // %[width] |
| 1794 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1795 | : "memory", "cc" |
| 1796 | #if defined(__SSE2__) |
| 1797 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1798 | #endif |
| 1799 | ); |
| 1800 | } |
| 1801 | |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1802 | void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf, |
| 1803 | const uint8* u_buf, |
| 1804 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 1805 | uint8* dst_rgb24, |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1806 | int width) { |
| 1807 | // fpic 32 bit gcc 4.2 on OSX runs out of GPR regs. |
| 1808 | #ifdef __APPLE__ |
| 1809 | asm volatile ( |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1810 | "movdqa %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n" |
| 1811 | "movdqa %[kShuffleMaskARGBToRGB24],%%xmm6 \n" |
| 1812 | :: [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0), |
| 1813 | [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)); |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1814 | #endif |
| 1815 | |
| 1816 | asm volatile ( |
| 1817 | #ifndef __APPLE__ |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1818 | "movdqa %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n" |
| 1819 | "movdqa %[kShuffleMaskARGBToRGB24],%%xmm6 \n" |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1820 | #endif |
| 1821 | "sub %[u_buf],%[v_buf] \n" |
| 1822 | "pxor %%xmm4,%%xmm4 \n" |
| 1823 | ".p2align 4 \n" |
| 1824 | "1: \n" |
| 1825 | READYUV422 |
| 1826 | YUVTORGB |
| 1827 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 1828 | "punpcklbw %%xmm2,%%xmm2 \n" |
| 1829 | "movdqa %%xmm0,%%xmm1 \n" |
| 1830 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 1831 | "punpckhwd %%xmm2,%%xmm1 \n" |
| 1832 | "pshufb %%xmm5,%%xmm0 \n" |
| 1833 | "pshufb %%xmm6,%%xmm1 \n" |
| 1834 | "palignr $0xc,%%xmm0,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1835 | "movq %%xmm0,(%[dst_rgb24]) \n" |
| 1836 | "movdqu %%xmm1,0x8(%[dst_rgb24]) \n" |
| 1837 | "lea 0x18(%[dst_rgb24]),%[dst_rgb24] \n" |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1838 | "sub $0x8,%[width] \n" |
| 1839 | "jg 1b \n" |
| 1840 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 1841 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 1842 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1843 | [dst_rgb24]"+r"(dst_rgb24), // %[dst_rgb24] |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1844 | [width]"+rm"(width) // %[width] |
| 1845 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) |
| 1846 | #ifndef __APPLE__ |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1847 | , [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0), |
| 1848 | [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24) |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1849 | #endif |
| 1850 | : "memory", "cc" |
| 1851 | #if defined(__SSE2__) |
| 1852 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
| 1853 | #endif |
| 1854 | ); |
| 1855 | } |
| 1856 | |
| 1857 | void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf, |
| 1858 | const uint8* u_buf, |
| 1859 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 1860 | uint8* dst_raw, |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1861 | int width) { |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1862 | // fpic 32 bit gcc 4.2 on OSX runs out of GPR regs. |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1863 | #ifdef __APPLE__ |
| 1864 | asm volatile ( |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1865 | "movdqa %[kShuffleMaskARGBToRAW_0],%%xmm5 \n" |
| 1866 | "movdqa %[kShuffleMaskARGBToRAW],%%xmm6 \n" |
| 1867 | :: [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0), |
| 1868 | [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)); |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1869 | #endif |
| 1870 | |
| 1871 | asm volatile ( |
| 1872 | #ifndef __APPLE__ |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1873 | "movdqa %[kShuffleMaskARGBToRAW_0],%%xmm5 \n" |
| 1874 | "movdqa %[kShuffleMaskARGBToRAW],%%xmm6 \n" |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1875 | #endif |
| 1876 | "sub %[u_buf],%[v_buf] \n" |
| 1877 | "pxor %%xmm4,%%xmm4 \n" |
| 1878 | ".p2align 4 \n" |
| 1879 | "1: \n" |
| 1880 | READYUV422 |
| 1881 | YUVTORGB |
| 1882 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 1883 | "punpcklbw %%xmm2,%%xmm2 \n" |
| 1884 | "movdqa %%xmm0,%%xmm1 \n" |
| 1885 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 1886 | "punpckhwd %%xmm2,%%xmm1 \n" |
| 1887 | "pshufb %%xmm5,%%xmm0 \n" |
| 1888 | "pshufb %%xmm6,%%xmm1 \n" |
| 1889 | "palignr $0xc,%%xmm0,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1890 | "movq %%xmm0,(%[dst_raw]) \n" |
| 1891 | "movdqu %%xmm1,0x8(%[dst_raw]) \n" |
| 1892 | "lea 0x18(%[dst_raw]),%[dst_raw] \n" |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1893 | "sub $0x8,%[width] \n" |
| 1894 | "jg 1b \n" |
| 1895 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 1896 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 1897 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1898 | [dst_raw]"+r"(dst_raw), // %[dst_raw] |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1899 | [width]"+rm"(width) // %[width] |
| 1900 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) |
| 1901 | #ifndef __APPLE__ |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1902 | , [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0), |
| 1903 | [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW) |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1904 | #endif |
| 1905 | : "memory", "cc" |
| 1906 | #if defined(__SSE2__) |
fbarchard@google.com | 4de0c43 | 2012-10-11 01:25:46 +0000 | [diff] [blame] | 1907 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 1908 | #endif |
| 1909 | ); |
| 1910 | } |
| 1911 | |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1912 | void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf, |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 1913 | const uint8* u_buf, |
| 1914 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 1915 | uint8* dst_argb, |
fbarchard@google.com | dbcabea | 2012-10-29 21:20:25 +0000 | [diff] [blame] | 1916 | int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 1917 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1918 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1919 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 1920 | "pxor %%xmm4,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 1921 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 1922 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1923 | READYUV422 |
| 1924 | YUVTORGB |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1925 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 1926 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 1927 | "movdqa %%xmm0,%%xmm1 \n" |
| 1928 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 1929 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1930 | "movdqa %%xmm0,(%[dst_argb]) \n" |
| 1931 | "movdqa %%xmm1,0x10(%[dst_argb]) \n" |
| 1932 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1933 | "sub $0x8,%[width] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1934 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1935 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 1936 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 1937 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1938 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1939 | [width]"+rm"(width) // %[width] |
| 1940 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1941 | : "memory", "cc" |
| 1942 | #if defined(__SSE2__) |
| 1943 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1944 | #endif |
| 1945 | ); |
| 1946 | } |
| 1947 | |
| 1948 | void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf, |
| 1949 | const uint8* u_buf, |
| 1950 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 1951 | uint8* dst_argb, |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1952 | int width) { |
| 1953 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1954 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1955 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 1956 | "pxor %%xmm4,%%xmm4 \n" |
| 1957 | ".p2align 4 \n" |
| 1958 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 1959 | READYUV411 |
| 1960 | YUVTORGB |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1961 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 1962 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 1963 | "movdqa %%xmm0,%%xmm1 \n" |
| 1964 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 1965 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1966 | "movdqa %%xmm0,(%[dst_argb]) \n" |
| 1967 | "movdqa %%xmm1,0x10(%[dst_argb]) \n" |
| 1968 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1969 | "sub $0x8,%[width] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 1970 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1971 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 1972 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 1973 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 1974 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1975 | [width]"+rm"(width) // %[width] |
| 1976 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
| 1977 | : "memory", "cc" |
| 1978 | #if defined(__SSE2__) |
| 1979 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 1980 | #endif |
| 1981 | ); |
| 1982 | } |
| 1983 | |
| 1984 | void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf, |
| 1985 | const uint8* uv_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 1986 | uint8* dst_argb, |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 1987 | int width) { |
| 1988 | asm volatile ( |
| 1989 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 1990 | "pxor %%xmm4,%%xmm4 \n" |
| 1991 | ".p2align 4 \n" |
| 1992 | "1: \n" |
| 1993 | READNV12 |
| 1994 | YUVTORGB |
| 1995 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 1996 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 1997 | "movdqa %%xmm0,%%xmm1 \n" |
| 1998 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 1999 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2000 | "movdqa %%xmm0,(%[dst_argb]) \n" |
| 2001 | "movdqa %%xmm1,0x10(%[dst_argb]) \n" |
| 2002 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2003 | "sub $0x8,%[width] \n" |
| 2004 | "jg 1b \n" |
| 2005 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2006 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2007 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2008 | [width]"+rm"(width) // %[width] |
| 2009 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
| 2010 | : "memory", "cc" |
| 2011 | #if defined(__SSE2__) |
| 2012 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2013 | #endif |
| 2014 | ); |
| 2015 | } |
| 2016 | |
| 2017 | void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf, |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2018 | const uint8* uv_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2019 | uint8* dst_argb, |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2020 | int width) { |
| 2021 | asm volatile ( |
| 2022 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2023 | "pxor %%xmm4,%%xmm4 \n" |
| 2024 | ".p2align 4 \n" |
| 2025 | "1: \n" |
| 2026 | READNV12 |
| 2027 | YVUTORGB |
| 2028 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2029 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 2030 | "movdqa %%xmm0,%%xmm1 \n" |
| 2031 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 2032 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2033 | "movdqa %%xmm0,(%[dst_argb]) \n" |
| 2034 | "movdqa %%xmm1,0x10(%[dst_argb]) \n" |
| 2035 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2036 | "sub $0x8,%[width] \n" |
| 2037 | "jg 1b \n" |
| 2038 | : [y_buf]"+r"(y_buf), // %[y_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2039 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
| 2040 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2041 | [width]"+rm"(width) // %[width] |
| 2042 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2043 | : "memory", "cc" |
| 2044 | #if defined(__SSE2__) |
| 2045 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2046 | #endif |
| 2047 | ); |
| 2048 | } |
| 2049 | |
| 2050 | void OMITFP I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf, |
| 2051 | const uint8* u_buf, |
| 2052 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2053 | uint8* dst_argb, |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2054 | int width) { |
| 2055 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2056 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2057 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2058 | "pxor %%xmm4,%%xmm4 \n" |
| 2059 | ".p2align 4 \n" |
| 2060 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2061 | READYUV444 |
| 2062 | YUVTORGB |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2063 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2064 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 2065 | "movdqa %%xmm0,%%xmm1 \n" |
| 2066 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 2067 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2068 | "movdqu %%xmm0,(%[dst_argb]) \n" |
| 2069 | "movdqu %%xmm1,0x10(%[dst_argb]) \n" |
| 2070 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2071 | "sub $0x8,%[width] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2072 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2073 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2074 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2075 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2076 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2077 | [width]"+rm"(width) // %[width] |
| 2078 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2079 | : "memory", "cc" |
| 2080 | #if defined(__SSE2__) |
| 2081 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2082 | #endif |
| 2083 | ); |
| 2084 | } |
| 2085 | |
| 2086 | void OMITFP I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf, |
| 2087 | const uint8* u_buf, |
| 2088 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2089 | uint8* dst_argb, |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2090 | int width) { |
| 2091 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2092 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2093 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2094 | "pxor %%xmm4,%%xmm4 \n" |
| 2095 | ".p2align 4 \n" |
| 2096 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2097 | READYUV422 |
| 2098 | YUVTORGB |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2099 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2100 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 2101 | "movdqa %%xmm0,%%xmm1 \n" |
| 2102 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 2103 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2104 | "movdqu %%xmm0,(%[dst_argb]) \n" |
| 2105 | "movdqu %%xmm1,0x10(%[dst_argb]) \n" |
| 2106 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2107 | "sub $0x8,%[width] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2108 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2109 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2110 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2111 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2112 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2113 | [width]"+rm"(width) // %[width] |
| 2114 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2115 | : "memory", "cc" |
| 2116 | #if defined(__SSE2__) |
| 2117 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2118 | #endif |
| 2119 | ); |
| 2120 | } |
| 2121 | |
| 2122 | void OMITFP I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf, |
| 2123 | const uint8* u_buf, |
| 2124 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2125 | uint8* dst_argb, |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2126 | int width) { |
| 2127 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2128 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2129 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2130 | "pxor %%xmm4,%%xmm4 \n" |
| 2131 | ".p2align 4 \n" |
| 2132 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2133 | READYUV411 |
| 2134 | YUVTORGB |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2135 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2136 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 2137 | "movdqa %%xmm0,%%xmm1 \n" |
| 2138 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 2139 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2140 | "movdqu %%xmm0,(%[dst_argb]) \n" |
| 2141 | "movdqu %%xmm1,0x10(%[dst_argb]) \n" |
| 2142 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2143 | "sub $0x8,%[width] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2144 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2145 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2146 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2147 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2148 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2149 | [width]"+rm"(width) // %[width] |
| 2150 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
| 2151 | : "memory", "cc" |
| 2152 | #if defined(__SSE2__) |
| 2153 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2154 | #endif |
| 2155 | ); |
| 2156 | } |
| 2157 | |
| 2158 | void OMITFP NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf, |
| 2159 | const uint8* uv_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2160 | uint8* dst_argb, |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2161 | int width) { |
| 2162 | asm volatile ( |
| 2163 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2164 | "pxor %%xmm4,%%xmm4 \n" |
| 2165 | ".p2align 4 \n" |
| 2166 | "1: \n" |
| 2167 | READNV12 |
| 2168 | YUVTORGB |
| 2169 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2170 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 2171 | "movdqa %%xmm0,%%xmm1 \n" |
| 2172 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 2173 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2174 | "movdqu %%xmm0,(%[dst_argb]) \n" |
| 2175 | "movdqu %%xmm1,0x10(%[dst_argb]) \n" |
| 2176 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2177 | "sub $0x8,%[width] \n" |
| 2178 | "jg 1b \n" |
| 2179 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2180 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2181 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2182 | [width]"+rm"(width) // %[width] |
| 2183 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
| 2184 | : "memory", "cc" |
| 2185 | #if defined(__SSE2__) |
| 2186 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2187 | #endif |
| 2188 | ); |
| 2189 | } |
| 2190 | |
| 2191 | void OMITFP NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf, |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2192 | const uint8* uv_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2193 | uint8* dst_argb, |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2194 | int width) { |
| 2195 | asm volatile ( |
| 2196 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2197 | "pxor %%xmm4,%%xmm4 \n" |
| 2198 | ".p2align 4 \n" |
| 2199 | "1: \n" |
| 2200 | READNV12 |
| 2201 | YVUTORGB |
| 2202 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2203 | "punpcklbw %%xmm5,%%xmm2 \n" |
| 2204 | "movdqa %%xmm0,%%xmm1 \n" |
| 2205 | "punpcklwd %%xmm2,%%xmm0 \n" |
| 2206 | "punpckhwd %%xmm2,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2207 | "movdqu %%xmm0,(%[dst_argb]) \n" |
| 2208 | "movdqu %%xmm1,0x10(%[dst_argb]) \n" |
| 2209 | "lea 0x20(%[dst_argb]),%[dst_argb] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2210 | "sub $0x8,%[width] \n" |
| 2211 | "jg 1b \n" |
| 2212 | : [y_buf]"+r"(y_buf), // %[y_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2213 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
| 2214 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2215 | [width]"+rm"(width) // %[width] |
| 2216 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2217 | : "memory", "cc" |
| 2218 | #if defined(__SSE2__) |
| 2219 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2220 | #endif |
| 2221 | ); |
| 2222 | } |
| 2223 | |
| 2224 | void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf, |
| 2225 | const uint8* u_buf, |
| 2226 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2227 | uint8* dst_bgra, |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2228 | int width) { |
| 2229 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2230 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2231 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2232 | "pxor %%xmm4,%%xmm4 \n" |
| 2233 | ".p2align 4 \n" |
| 2234 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2235 | READYUV422 |
| 2236 | YUVTORGB |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2237 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2238 | "punpcklbw %%xmm0,%%xmm1 \n" |
| 2239 | "punpcklbw %%xmm2,%%xmm5 \n" |
| 2240 | "movdqa %%xmm5,%%xmm0 \n" |
| 2241 | "punpcklwd %%xmm1,%%xmm5 \n" |
| 2242 | "punpckhwd %%xmm1,%%xmm0 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2243 | "movdqa %%xmm5,(%[dst_bgra]) \n" |
| 2244 | "movdqa %%xmm0,0x10(%[dst_bgra]) \n" |
| 2245 | "lea 0x20(%[dst_bgra]),%[dst_bgra] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2246 | "sub $0x8,%[width] \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2247 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2248 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2249 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2250 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2251 | [dst_bgra]"+r"(dst_bgra), // %[dst_bgra] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2252 | [width]"+rm"(width) // %[width] |
| 2253 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2254 | : "memory", "cc" |
| 2255 | #if defined(__SSE2__) |
| 2256 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2257 | #endif |
| 2258 | ); |
| 2259 | } |
| 2260 | |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2261 | void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf, |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2262 | const uint8* u_buf, |
| 2263 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2264 | uint8* dst_abgr, |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2265 | int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2266 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2267 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2268 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2269 | "pxor %%xmm4,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2270 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2271 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2272 | READYUV422 |
| 2273 | YUVTORGB |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2274 | "punpcklbw %%xmm1,%%xmm2 \n" |
| 2275 | "punpcklbw %%xmm5,%%xmm0 \n" |
| 2276 | "movdqa %%xmm2,%%xmm1 \n" |
| 2277 | "punpcklwd %%xmm0,%%xmm2 \n" |
| 2278 | "punpckhwd %%xmm0,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2279 | "movdqa %%xmm2,(%[dst_abgr]) \n" |
| 2280 | "movdqa %%xmm1,0x10(%[dst_abgr]) \n" |
| 2281 | "lea 0x20(%[dst_abgr]),%[dst_abgr] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2282 | "sub $0x8,%[width] \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2283 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2284 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2285 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2286 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2287 | [dst_abgr]"+r"(dst_abgr), // %[dst_abgr] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2288 | [width]"+rm"(width) // %[width] |
| 2289 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2290 | : "memory", "cc" |
| 2291 | #if defined(__SSE2__) |
| 2292 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2293 | #endif |
| 2294 | ); |
| 2295 | } |
| 2296 | |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2297 | void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf, |
| 2298 | const uint8* u_buf, |
| 2299 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2300 | uint8* dst_rgba, |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2301 | int width) { |
| 2302 | asm volatile ( |
| 2303 | "sub %[u_buf],%[v_buf] \n" |
| 2304 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2305 | "pxor %%xmm4,%%xmm4 \n" |
| 2306 | ".p2align 4 \n" |
| 2307 | "1: \n" |
| 2308 | READYUV422 |
| 2309 | YUVTORGB |
| 2310 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2311 | "punpcklbw %%xmm2,%%xmm1 \n" |
| 2312 | "punpcklbw %%xmm0,%%xmm5 \n" |
| 2313 | "movdqa %%xmm5,%%xmm0 \n" |
| 2314 | "punpcklwd %%xmm1,%%xmm5 \n" |
| 2315 | "punpckhwd %%xmm1,%%xmm0 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2316 | "movdqa %%xmm5,(%[dst_rgba]) \n" |
| 2317 | "movdqa %%xmm0,0x10(%[dst_rgba]) \n" |
| 2318 | "lea 0x20(%[dst_rgba]),%[dst_rgba] \n" |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2319 | "sub $0x8,%[width] \n" |
| 2320 | "jg 1b \n" |
| 2321 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2322 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2323 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2324 | [dst_rgba]"+r"(dst_rgba), // %[dst_rgba] |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2325 | [width]"+rm"(width) // %[width] |
| 2326 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
| 2327 | : "memory", "cc" |
| 2328 | #if defined(__SSE2__) |
| 2329 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2330 | #endif |
| 2331 | ); |
| 2332 | } |
| 2333 | |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2334 | void OMITFP I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf, |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2335 | const uint8* u_buf, |
| 2336 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2337 | uint8* dst_bgra, |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2338 | int width) { |
| 2339 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2340 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2341 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2342 | "pxor %%xmm4,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2343 | ".p2align 4 \n" |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2344 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2345 | READYUV422 |
| 2346 | YUVTORGB |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2347 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2348 | "punpcklbw %%xmm0,%%xmm1 \n" |
| 2349 | "punpcklbw %%xmm2,%%xmm5 \n" |
| 2350 | "movdqa %%xmm5,%%xmm0 \n" |
| 2351 | "punpcklwd %%xmm1,%%xmm5 \n" |
| 2352 | "punpckhwd %%xmm1,%%xmm0 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2353 | "movdqu %%xmm5,(%[dst_bgra]) \n" |
| 2354 | "movdqu %%xmm0,0x10(%[dst_bgra]) \n" |
| 2355 | "lea 0x20(%[dst_bgra]),%[dst_bgra] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2356 | "sub $0x8,%[width] \n" |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2357 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2358 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2359 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2360 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2361 | [dst_bgra]"+r"(dst_bgra), // %[dst_bgra] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2362 | [width]"+rm"(width) // %[width] |
| 2363 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2364 | : "memory", "cc" |
| 2365 | #if defined(__SSE2__) |
| 2366 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2367 | #endif |
| 2368 | ); |
| 2369 | } |
| 2370 | |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2371 | void OMITFP I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf, |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2372 | const uint8* u_buf, |
| 2373 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2374 | uint8* dst_abgr, |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2375 | int width) { |
| 2376 | asm volatile ( |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2377 | "sub %[u_buf],%[v_buf] \n" |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2378 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2379 | "pxor %%xmm4,%%xmm4 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2380 | ".p2align 4 \n" |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2381 | "1: \n" |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2382 | READYUV422 |
| 2383 | YUVTORGB |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2384 | "punpcklbw %%xmm1,%%xmm2 \n" |
| 2385 | "punpcklbw %%xmm5,%%xmm0 \n" |
| 2386 | "movdqa %%xmm2,%%xmm1 \n" |
| 2387 | "punpcklwd %%xmm0,%%xmm2 \n" |
| 2388 | "punpckhwd %%xmm0,%%xmm1 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2389 | "movdqu %%xmm2,(%[dst_abgr]) \n" |
| 2390 | "movdqu %%xmm1,0x10(%[dst_abgr]) \n" |
| 2391 | "lea 0x20(%[dst_abgr]),%[dst_abgr] \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2392 | "sub $0x8,%[width] \n" |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2393 | "jg 1b \n" |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2394 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2395 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2396 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2397 | [dst_abgr]"+r"(dst_abgr), // %[dst_abgr] |
fbarchard@google.com | 2d9fe08 | 2012-06-05 22:11:34 +0000 | [diff] [blame] | 2398 | [width]"+rm"(width) // %[width] |
| 2399 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
fbarchard@google.com | 952a507 | 2012-03-30 18:10:50 +0000 | [diff] [blame] | 2400 | : "memory", "cc" |
| 2401 | #if defined(__SSE2__) |
| 2402 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2403 | #endif |
| 2404 | ); |
| 2405 | } |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2406 | |
| 2407 | void OMITFP I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf, |
| 2408 | const uint8* u_buf, |
| 2409 | const uint8* v_buf, |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2410 | uint8* dst_rgba, |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2411 | int width) { |
| 2412 | asm volatile ( |
| 2413 | "sub %[u_buf],%[v_buf] \n" |
| 2414 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2415 | "pxor %%xmm4,%%xmm4 \n" |
| 2416 | ".p2align 4 \n" |
| 2417 | "1: \n" |
| 2418 | READYUV422 |
| 2419 | YUVTORGB |
| 2420 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2421 | "punpcklbw %%xmm2,%%xmm1 \n" |
| 2422 | "punpcklbw %%xmm0,%%xmm5 \n" |
| 2423 | "movdqa %%xmm5,%%xmm0 \n" |
| 2424 | "punpcklwd %%xmm1,%%xmm5 \n" |
| 2425 | "punpckhwd %%xmm1,%%xmm0 \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2426 | "movdqa %%xmm5,(%[dst_rgba]) \n" |
| 2427 | "movdqa %%xmm0,0x10(%[dst_rgba]) \n" |
| 2428 | "lea 0x20(%[dst_rgba]),%[dst_rgba] \n" |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2429 | "sub $0x8,%[width] \n" |
| 2430 | "jg 1b \n" |
| 2431 | : [y_buf]"+r"(y_buf), // %[y_buf] |
| 2432 | [u_buf]"+r"(u_buf), // %[u_buf] |
| 2433 | [v_buf]"+r"(v_buf), // %[v_buf] |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2434 | [dst_rgba]"+r"(dst_rgba), // %[dst_rgba] |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 2435 | [width]"+rm"(width) // %[width] |
| 2436 | : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] |
| 2437 | : "memory", "cc" |
| 2438 | #if defined(__SSE2__) |
| 2439 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 2440 | #endif |
| 2441 | ); |
| 2442 | } |
| 2443 | |
fbarchard@google.com | e214fe3 | 2012-06-04 23:47:11 +0000 | [diff] [blame] | 2444 | #endif // HAS_I422TOARGBROW_SSSE3 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2445 | |
| 2446 | #ifdef HAS_YTOARGBROW_SSE2 |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2447 | void YToARGBRow_SSE2(const uint8* y_buf, |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2448 | uint8* dst_argb, |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2449 | int width) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2450 | asm volatile ( |
fbarchard@google.com | 30859f7 | 2012-11-02 09:51:29 +0000 | [diff] [blame] | 2451 | "pxor %%xmm5,%%xmm5 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2452 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 2453 | "pslld $0x18,%%xmm4 \n" |
fbarchard@google.com | 30859f7 | 2012-11-02 09:51:29 +0000 | [diff] [blame] | 2454 | "mov $0x00100010,%%eax \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2455 | "movd %%eax,%%xmm3 \n" |
| 2456 | "pshufd $0x0,%%xmm3,%%xmm3 \n" |
fbarchard@google.com | 30859f7 | 2012-11-02 09:51:29 +0000 | [diff] [blame] | 2457 | "mov $0x004a004a,%%eax \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2458 | "movd %%eax,%%xmm2 \n" |
| 2459 | "pshufd $0x0,%%xmm2,%%xmm2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2460 | ".p2align 4 \n" |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 2461 | "1: \n" |
fbarchard@google.com | 8b9759c | 2011-12-14 04:17:39 +0000 | [diff] [blame] | 2462 | // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2463 | "movq (%0),%%xmm0 \n" |
| 2464 | "lea 0x8(%0),%0 \n" |
fbarchard@google.com | 30859f7 | 2012-11-02 09:51:29 +0000 | [diff] [blame] | 2465 | "punpcklbw %%xmm5,%%xmm0 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2466 | "psubusw %%xmm3,%%xmm0 \n" |
fbarchard@google.com | 30859f7 | 2012-11-02 09:51:29 +0000 | [diff] [blame] | 2467 | "pmullw %%xmm2,%%xmm0 \n" |
| 2468 | "psrlw $6, %%xmm0 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2469 | "packuswb %%xmm0,%%xmm0 \n" |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 2470 | |
fbarchard@google.com | 8b9759c | 2011-12-14 04:17:39 +0000 | [diff] [blame] | 2471 | // Step 2: Weave into ARGB |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2472 | "punpcklbw %%xmm0,%%xmm0 \n" |
| 2473 | "movdqa %%xmm0,%%xmm1 \n" |
| 2474 | "punpcklwd %%xmm0,%%xmm0 \n" |
| 2475 | "punpckhwd %%xmm1,%%xmm1 \n" |
| 2476 | "por %%xmm4,%%xmm0 \n" |
| 2477 | "por %%xmm4,%%xmm1 \n" |
| 2478 | "movdqa %%xmm0,(%1) \n" |
| 2479 | "movdqa %%xmm1,16(%1) \n" |
| 2480 | "lea 32(%1),%1 \n" |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 2481 | |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2482 | "sub $0x8,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2483 | "jg 1b \n" |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 2484 | : "+r"(y_buf), // %0 |
| 2485 | "+r"(dst_argb), // %1 |
| 2486 | "+rm"(width) // %2 |
fbarchard@google.com | 8b9759c | 2011-12-14 04:17:39 +0000 | [diff] [blame] | 2487 | : |
| 2488 | : "memory", "cc", "eax" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 2489 | #if defined(__SSE2__) |
fbarchard@google.com | 8b9759c | 2011-12-14 04:17:39 +0000 | [diff] [blame] | 2490 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" |
fbarchard@google.com | b614976 | 2011-11-07 21:58:52 +0000 | [diff] [blame] | 2491 | #endif |
fbarchard@google.com | 228bdc2 | 2011-11-15 21:58:26 +0000 | [diff] [blame] | 2492 | ); |
mikhal@webrtc.org | 120d5e7 | 2011-10-07 17:57:17 +0000 | [diff] [blame] | 2493 | } |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2494 | #endif // HAS_YTOARGBROW_SSE2 |
mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 2495 | |
fbarchard@google.com | 42831e0 | 2012-01-21 02:54:17 +0000 | [diff] [blame] | 2496 | #ifdef HAS_MIRRORROW_SSSE3 |
fbarchard@google.com | 12d0483 | 2011-11-21 23:54:38 +0000 | [diff] [blame] | 2497 | // Shuffle table for reversing the bytes. |
fbarchard@google.com | 42831e0 | 2012-01-21 02:54:17 +0000 | [diff] [blame] | 2498 | CONST uvec8 kShuffleMirror = { |
fbarchard@google.com | 12d0483 | 2011-11-21 23:54:38 +0000 | [diff] [blame] | 2499 | 15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u |
| 2500 | }; |
| 2501 | |
fbarchard@google.com | 42831e0 | 2012-01-21 02:54:17 +0000 | [diff] [blame] | 2502 | void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) { |
fbarchard@google.com | 12d0483 | 2011-11-21 23:54:38 +0000 | [diff] [blame] | 2503 | intptr_t temp_width = static_cast<intptr_t>(width); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2504 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2505 | "movdqa %3,%%xmm5 \n" |
| 2506 | "lea -0x10(%0),%0 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2507 | ".p2align 4 \n" |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 2508 | "1: \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2509 | "movdqa (%0,%2),%%xmm0 \n" |
| 2510 | "pshufb %%xmm5,%%xmm0 \n" |
| 2511 | "sub $0x10,%2 \n" |
| 2512 | "movdqa %%xmm0,(%1) \n" |
| 2513 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2514 | "jg 1b \n" |
fbarchard@google.com | 12d0483 | 2011-11-21 23:54:38 +0000 | [diff] [blame] | 2515 | : "+r"(src), // %0 |
| 2516 | "+r"(dst), // %1 |
| 2517 | "+r"(temp_width) // %2 |
fbarchard@google.com | 42831e0 | 2012-01-21 02:54:17 +0000 | [diff] [blame] | 2518 | : "m"(kShuffleMirror) // %3 |
fbarchard@google.com | 12d0483 | 2011-11-21 23:54:38 +0000 | [diff] [blame] | 2519 | : "memory", "cc" |
| 2520 | #if defined(__SSE2__) |
| 2521 | , "xmm0", "xmm5" |
| 2522 | #endif |
| 2523 | ); |
| 2524 | } |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2525 | #endif // HAS_MIRRORROW_SSSE3 |
fbarchard@google.com | 12d0483 | 2011-11-21 23:54:38 +0000 | [diff] [blame] | 2526 | |
fbarchard@google.com | 42831e0 | 2012-01-21 02:54:17 +0000 | [diff] [blame] | 2527 | #ifdef HAS_MIRRORROW_SSE2 |
fbarchard@google.com | 42831e0 | 2012-01-21 02:54:17 +0000 | [diff] [blame] | 2528 | void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) { |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 2529 | intptr_t temp_width = static_cast<intptr_t>(width); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2530 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2531 | "lea -0x10(%0),%0 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2532 | ".p2align 4 \n" |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 2533 | "1: \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2534 | "movdqu (%0,%2),%%xmm0 \n" |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2535 | "movdqa %%xmm0,%%xmm1 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2536 | "psllw $0x8,%%xmm0 \n" |
| 2537 | "psrlw $0x8,%%xmm1 \n" |
| 2538 | "por %%xmm1,%%xmm0 \n" |
| 2539 | "pshuflw $0x1b,%%xmm0,%%xmm0 \n" |
| 2540 | "pshufhw $0x1b,%%xmm0,%%xmm0 \n" |
| 2541 | "pshufd $0x4e,%%xmm0,%%xmm0 \n" |
| 2542 | "sub $0x10,%2 \n" |
| 2543 | "movdqu %%xmm0,(%1) \n" |
| 2544 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2545 | "jg 1b \n" |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 2546 | : "+r"(src), // %0 |
| 2547 | "+r"(dst), // %1 |
| 2548 | "+r"(temp_width) // %2 |
| 2549 | : |
| 2550 | : "memory", "cc" |
| 2551 | #if defined(__SSE2__) |
| 2552 | , "xmm0", "xmm1" |
| 2553 | #endif |
| 2554 | ); |
| 2555 | } |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2556 | #endif // HAS_MIRRORROW_SSE2 |
fbarchard@google.com | 373cdbd | 2011-12-14 21:10:07 +0000 | [diff] [blame] | 2557 | |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2558 | #ifdef HAS_MIRRORROW_UV_SSSE3 |
| 2559 | // Shuffle table for reversing the bytes of UV channels. |
| 2560 | CONST uvec8 kShuffleMirrorUV = { |
| 2561 | 14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u |
| 2562 | }; |
fbarchard@google.com | bdf7cb5 | 2012-11-05 23:40:11 +0000 | [diff] [blame] | 2563 | void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v, |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2564 | int width) { |
| 2565 | intptr_t temp_width = static_cast<intptr_t>(width); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2566 | asm volatile ( |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2567 | "movdqa %4,%%xmm1 \n" |
| 2568 | "lea -16(%0,%3,2),%0 \n" |
| 2569 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2570 | ".p2align 4 \n" |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2571 | "1: \n" |
| 2572 | "movdqa (%0),%%xmm0 \n" |
| 2573 | "lea -16(%0),%0 \n" |
| 2574 | "pshufb %%xmm1,%%xmm0 \n" |
| 2575 | "sub $8,%3 \n" |
| 2576 | "movlpd %%xmm0,(%1) \n" |
| 2577 | "movhpd %%xmm0,(%1,%2) \n" |
| 2578 | "lea 8(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2579 | "jg 1b \n" |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2580 | : "+r"(src), // %0 |
| 2581 | "+r"(dst_u), // %1 |
| 2582 | "+r"(dst_v), // %2 |
| 2583 | "+r"(temp_width) // %3 |
fbarchard@google.com | 4d2cfd3 | 2012-06-25 20:09:31 +0000 | [diff] [blame] | 2584 | : "m"(kShuffleMirrorUV) // %4 |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2585 | : "memory", "cc" |
| 2586 | #if defined(__SSE2__) |
| 2587 | , "xmm0", "xmm1" |
| 2588 | #endif |
| 2589 | ); |
| 2590 | } |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2591 | #endif // HAS_MIRRORROW_UV_SSSE3 |
fbarchard@google.com | 16a9664 | 2012-03-02 22:38:09 +0000 | [diff] [blame] | 2592 | |
fbarchard@google.com | 27d42c7 | 2012-06-22 23:57:26 +0000 | [diff] [blame] | 2593 | #ifdef HAS_ARGBMIRRORROW_SSSE3 |
| 2594 | // Shuffle table for reversing the bytes. |
fbarchard@google.com | 4d2cfd3 | 2012-06-25 20:09:31 +0000 | [diff] [blame] | 2595 | CONST uvec8 kARGBShuffleMirror = { |
fbarchard@google.com | 27d42c7 | 2012-06-22 23:57:26 +0000 | [diff] [blame] | 2596 | 12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u |
| 2597 | }; |
| 2598 | |
| 2599 | void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) { |
| 2600 | intptr_t temp_width = static_cast<intptr_t>(width); |
| 2601 | asm volatile ( |
| 2602 | "movdqa %3,%%xmm5 \n" |
| 2603 | "lea -0x10(%0),%0 \n" |
| 2604 | ".p2align 4 \n" |
| 2605 | "1: \n" |
| 2606 | "movdqa (%0,%2,4),%%xmm0 \n" |
| 2607 | "pshufb %%xmm5,%%xmm0 \n" |
| 2608 | "sub $0x4,%2 \n" |
| 2609 | "movdqa %%xmm0,(%1) \n" |
| 2610 | "lea 0x10(%1),%1 \n" |
| 2611 | "jg 1b \n" |
| 2612 | : "+r"(src), // %0 |
| 2613 | "+r"(dst), // %1 |
| 2614 | "+r"(temp_width) // %2 |
fbarchard@google.com | 4d2cfd3 | 2012-06-25 20:09:31 +0000 | [diff] [blame] | 2615 | : "m"(kARGBShuffleMirror) // %3 |
fbarchard@google.com | 27d42c7 | 2012-06-22 23:57:26 +0000 | [diff] [blame] | 2616 | : "memory", "cc" |
| 2617 | #if defined(__SSE2__) |
| 2618 | , "xmm0", "xmm5" |
| 2619 | #endif |
| 2620 | ); |
| 2621 | } |
| 2622 | #endif // HAS_ARGBMIRRORROW_SSSE3 |
| 2623 | |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2624 | #ifdef HAS_SPLITUV_SSE2 |
| 2625 | void SplitUV_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2626 | asm volatile ( |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2627 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2628 | "psrlw $0x8,%%xmm5 \n" |
| 2629 | "sub %1,%2 \n" |
fbarchard@google.com | db694ed | 2012-10-17 21:54:04 +0000 | [diff] [blame] | 2630 | ".p2align 4 \n" |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2631 | "1: \n" |
| 2632 | "movdqa (%0),%%xmm0 \n" |
| 2633 | "movdqa 0x10(%0),%%xmm1 \n" |
| 2634 | "lea 0x20(%0),%0 \n" |
| 2635 | "movdqa %%xmm0,%%xmm2 \n" |
| 2636 | "movdqa %%xmm1,%%xmm3 \n" |
| 2637 | "pand %%xmm5,%%xmm0 \n" |
| 2638 | "pand %%xmm5,%%xmm1 \n" |
| 2639 | "packuswb %%xmm1,%%xmm0 \n" |
| 2640 | "psrlw $0x8,%%xmm2 \n" |
| 2641 | "psrlw $0x8,%%xmm3 \n" |
| 2642 | "packuswb %%xmm3,%%xmm2 \n" |
| 2643 | "movdqa %%xmm0,(%1) \n" |
| 2644 | "movdqa %%xmm2,(%1,%2) \n" |
| 2645 | "lea 0x10(%1),%1 \n" |
| 2646 | "sub $0x10,%3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2647 | "jg 1b \n" |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2648 | : "+r"(src_uv), // %0 |
| 2649 | "+r"(dst_u), // %1 |
| 2650 | "+r"(dst_v), // %2 |
| 2651 | "+r"(pix) // %3 |
| 2652 | : |
| 2653 | : "memory", "cc" |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2654 | #if defined(__SSE2__) |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2655 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2656 | #endif |
| 2657 | ); |
| 2658 | } |
fbarchard@google.com | db694ed | 2012-10-17 21:54:04 +0000 | [diff] [blame] | 2659 | |
| 2660 | void SplitUV_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, |
| 2661 | int pix) { |
| 2662 | asm volatile ( |
| 2663 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2664 | "psrlw $0x8,%%xmm5 \n" |
| 2665 | "sub %1,%2 \n" |
| 2666 | ".p2align 4 \n" |
| 2667 | "1: \n" |
| 2668 | "movdqu (%0),%%xmm0 \n" |
| 2669 | "movdqu 0x10(%0),%%xmm1 \n" |
| 2670 | "lea 0x20(%0),%0 \n" |
| 2671 | "movdqa %%xmm0,%%xmm2 \n" |
| 2672 | "movdqa %%xmm1,%%xmm3 \n" |
| 2673 | "pand %%xmm5,%%xmm0 \n" |
| 2674 | "pand %%xmm5,%%xmm1 \n" |
| 2675 | "packuswb %%xmm1,%%xmm0 \n" |
| 2676 | "psrlw $0x8,%%xmm2 \n" |
| 2677 | "psrlw $0x8,%%xmm3 \n" |
| 2678 | "packuswb %%xmm3,%%xmm2 \n" |
| 2679 | "movdqu %%xmm0,(%1) \n" |
| 2680 | "movdqu %%xmm2,(%1,%2) \n" |
| 2681 | "lea 0x10(%1),%1 \n" |
| 2682 | "sub $0x10,%3 \n" |
| 2683 | "jg 1b \n" |
| 2684 | : "+r"(src_uv), // %0 |
| 2685 | "+r"(dst_u), // %1 |
| 2686 | "+r"(dst_v), // %2 |
| 2687 | "+r"(pix) // %3 |
| 2688 | : |
| 2689 | : "memory", "cc" |
| 2690 | #if defined(__SSE2__) |
| 2691 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" |
| 2692 | #endif |
| 2693 | ); |
| 2694 | } |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2695 | #endif // HAS_SPLITUV_SSE2 |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 2696 | |
fbarchard@google.com | 1dafd44 | 2012-10-26 08:27:36 +0000 | [diff] [blame] | 2697 | #ifdef HAS_MERGEUV_SSE2 |
| 2698 | void MergeUV_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, |
| 2699 | int width) { |
| 2700 | asm volatile ( |
| 2701 | "sub %0,%1 \n" |
| 2702 | ".p2align 4 \n" |
| 2703 | "1: \n" |
| 2704 | "movdqa (%0),%%xmm0 \n" |
| 2705 | "movdqa (%0,%1,1),%%xmm1 \n" |
| 2706 | "lea 0x10(%0),%0 \n" |
| 2707 | "movdqa %%xmm0,%%xmm2 \n" |
| 2708 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2709 | "punpckhbw %%xmm1,%%xmm2 \n" |
| 2710 | "movdqa %%xmm0,(%2) \n" |
| 2711 | "movdqa %%xmm2,0x10(%2) \n" |
| 2712 | "lea 0x20(%2),%2 \n" |
| 2713 | "sub $0x10,%3 \n" |
| 2714 | "jg 1b \n" |
| 2715 | : "+r"(src_u), // %0 |
| 2716 | "+r"(src_v), // %1 |
| 2717 | "+r"(dst_uv), // %2 |
| 2718 | "+r"(width) // %3 |
| 2719 | : |
| 2720 | : "memory", "cc" |
| 2721 | #if defined(__SSE2__) |
| 2722 | , "xmm0", "xmm1", "xmm2" |
| 2723 | #endif |
| 2724 | ); |
| 2725 | } |
fbarchard@google.com | e0d8648 | 2012-10-27 19:07:55 +0000 | [diff] [blame] | 2726 | |
| 2727 | void MergeUV_Unaligned_SSE2(const uint8* src_u, const uint8* src_v, |
| 2728 | uint8* dst_uv, int width) { |
| 2729 | asm volatile ( |
| 2730 | "sub %0,%1 \n" |
| 2731 | ".p2align 4 \n" |
| 2732 | "1: \n" |
| 2733 | "movdqu (%0),%%xmm0 \n" |
| 2734 | "movdqu (%0,%1,1),%%xmm1 \n" |
| 2735 | "lea 0x10(%0),%0 \n" |
| 2736 | "movdqa %%xmm0,%%xmm2 \n" |
| 2737 | "punpcklbw %%xmm1,%%xmm0 \n" |
| 2738 | "punpckhbw %%xmm1,%%xmm2 \n" |
| 2739 | "movdqu %%xmm0,(%2) \n" |
| 2740 | "movdqu %%xmm2,0x10(%2) \n" |
| 2741 | "lea 0x20(%2),%2 \n" |
| 2742 | "sub $0x10,%3 \n" |
| 2743 | "jg 1b \n" |
| 2744 | : "+r"(src_u), // %0 |
| 2745 | "+r"(src_v), // %1 |
| 2746 | "+r"(dst_uv), // %2 |
| 2747 | "+r"(width) // %3 |
| 2748 | : |
| 2749 | : "memory", "cc" |
| 2750 | #if defined(__SSE2__) |
| 2751 | , "xmm0", "xmm1", "xmm2" |
| 2752 | #endif |
| 2753 | ); |
| 2754 | } |
fbarchard@google.com | 1dafd44 | 2012-10-26 08:27:36 +0000 | [diff] [blame] | 2755 | #endif // HAS_MERGEUV_SSE2 |
| 2756 | |
fbarchard@google.com | 19932f8 | 2012-02-16 22:19:14 +0000 | [diff] [blame] | 2757 | #ifdef HAS_COPYROW_SSE2 |
| 2758 | void CopyRow_SSE2(const uint8* src, uint8* dst, int count) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2759 | asm volatile ( |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2760 | "sub %0,%1 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 2761 | ".p2align 4 \n" |
fbarchard@google.com | 19932f8 | 2012-02-16 22:19:14 +0000 | [diff] [blame] | 2762 | "1: \n" |
| 2763 | "movdqa (%0),%%xmm0 \n" |
| 2764 | "movdqa 0x10(%0),%%xmm1 \n" |
| 2765 | "movdqa %%xmm0,(%0,%1) \n" |
| 2766 | "movdqa %%xmm1,0x10(%0,%1) \n" |
| 2767 | "lea 0x20(%0),%0 \n" |
| 2768 | "sub $0x20,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2769 | "jg 1b \n" |
fbarchard@google.com | 19932f8 | 2012-02-16 22:19:14 +0000 | [diff] [blame] | 2770 | : "+r"(src), // %0 |
| 2771 | "+r"(dst), // %1 |
| 2772 | "+r"(count) // %2 |
| 2773 | : |
| 2774 | : "memory", "cc" |
| 2775 | #if defined(__SSE2__) |
| 2776 | , "xmm0", "xmm1" |
| 2777 | #endif |
| 2778 | ); |
| 2779 | } |
| 2780 | #endif // HAS_COPYROW_SSE2 |
| 2781 | |
| 2782 | #ifdef HAS_COPYROW_X86 |
| 2783 | void CopyRow_X86(const uint8* src, uint8* dst, int width) { |
| 2784 | size_t width_tmp = static_cast<size_t>(width); |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2785 | asm volatile ( |
fbarchard@google.com | 19932f8 | 2012-02-16 22:19:14 +0000 | [diff] [blame] | 2786 | "shr $0x2,%2 \n" |
| 2787 | "rep movsl \n" |
| 2788 | : "+S"(src), // %0 |
| 2789 | "+D"(dst), // %1 |
| 2790 | "+c"(width_tmp) // %2 |
| 2791 | : |
| 2792 | : "memory", "cc" |
| 2793 | ); |
| 2794 | } |
fbarchard@google.com | 4c416e8 | 2012-06-05 15:48:07 +0000 | [diff] [blame] | 2795 | #endif // HAS_COPYROW_X86 |
fbarchard@google.com | 19932f8 | 2012-02-16 22:19:14 +0000 | [diff] [blame] | 2796 | |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 2797 | #ifdef HAS_SETROW_X86 |
| 2798 | void SetRow8_X86(uint8* dst, uint32 v32, int width) { |
| 2799 | size_t width_tmp = static_cast<size_t>(width); |
| 2800 | asm volatile ( |
| 2801 | "shr $0x2,%1 \n" |
| 2802 | "rep stosl \n" |
| 2803 | : "+D"(dst), // %0 |
| 2804 | "+c"(width_tmp) // %1 |
| 2805 | : "a"(v32) // %2 |
| 2806 | : "memory", "cc"); |
| 2807 | } |
| 2808 | |
| 2809 | void SetRows32_X86(uint8* dst, uint32 v32, int width, |
| 2810 | int dst_stride, int height) { |
| 2811 | for (int y = 0; y < height; ++y) { |
| 2812 | size_t width_tmp = static_cast<size_t>(width); |
| 2813 | uint32* d = reinterpret_cast<uint32*>(dst); |
| 2814 | asm volatile ( |
| 2815 | "rep stosl \n" |
| 2816 | : "+D"(d), // %0 |
| 2817 | "+c"(width_tmp) // %1 |
| 2818 | : "a"(v32) // %2 |
| 2819 | : "memory", "cc"); |
| 2820 | dst += dst_stride; |
| 2821 | } |
| 2822 | } |
| 2823 | #endif // HAS_SETROW_X86 |
| 2824 | |
fbarchard@google.com | b95dbf2 | 2012-02-11 01:18:30 +0000 | [diff] [blame] | 2825 | #ifdef HAS_YUY2TOYROW_SSE2 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2826 | void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2827 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2828 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2829 | "psrlw $0x8,%%xmm5 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2830 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2831 | "1: \n" |
| 2832 | "movdqa (%0),%%xmm0 \n" |
| 2833 | "movdqa 0x10(%0),%%xmm1 \n" |
| 2834 | "lea 0x20(%0),%0 \n" |
| 2835 | "pand %%xmm5,%%xmm0 \n" |
| 2836 | "pand %%xmm5,%%xmm1 \n" |
| 2837 | "packuswb %%xmm1,%%xmm0 \n" |
| 2838 | "movdqa %%xmm0,(%1) \n" |
| 2839 | "lea 0x10(%1),%1 \n" |
| 2840 | "sub $0x10,%2 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2841 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2842 | : "+r"(src_yuy2), // %0 |
| 2843 | "+r"(dst_y), // %1 |
| 2844 | "+r"(pix) // %2 |
| 2845 | : |
| 2846 | : "memory", "cc" |
| 2847 | #if defined(__SSE2__) |
| 2848 | , "xmm0", "xmm1", "xmm5" |
| 2849 | #endif |
| 2850 | ); |
| 2851 | } |
| 2852 | |
| 2853 | void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2, |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 2854 | uint8* dst_u, uint8* dst_v, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2855 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2856 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2857 | "psrlw $0x8,%%xmm5 \n" |
| 2858 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2859 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2860 | "1: \n" |
| 2861 | "movdqa (%0),%%xmm0 \n" |
| 2862 | "movdqa 0x10(%0),%%xmm1 \n" |
| 2863 | "movdqa (%0,%4,1),%%xmm2 \n" |
| 2864 | "movdqa 0x10(%0,%4,1),%%xmm3 \n" |
| 2865 | "lea 0x20(%0),%0 \n" |
| 2866 | "pavgb %%xmm2,%%xmm0 \n" |
| 2867 | "pavgb %%xmm3,%%xmm1 \n" |
| 2868 | "psrlw $0x8,%%xmm0 \n" |
| 2869 | "psrlw $0x8,%%xmm1 \n" |
| 2870 | "packuswb %%xmm1,%%xmm0 \n" |
| 2871 | "movdqa %%xmm0,%%xmm1 \n" |
| 2872 | "pand %%xmm5,%%xmm0 \n" |
| 2873 | "packuswb %%xmm0,%%xmm0 \n" |
| 2874 | "psrlw $0x8,%%xmm1 \n" |
| 2875 | "packuswb %%xmm1,%%xmm1 \n" |
| 2876 | "movq %%xmm0,(%1) \n" |
| 2877 | "movq %%xmm1,(%1,%2) \n" |
| 2878 | "lea 0x8(%1),%1 \n" |
| 2879 | "sub $0x10,%3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2880 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2881 | : "+r"(src_yuy2), // %0 |
| 2882 | "+r"(dst_u), // %1 |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 2883 | "+r"(dst_v), // %2 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2884 | "+r"(pix) // %3 |
| 2885 | : "r"(static_cast<intptr_t>(stride_yuy2)) // %4 |
| 2886 | : "memory", "cc" |
| 2887 | #if defined(__SSE2__) |
| 2888 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" |
| 2889 | #endif |
| 2890 | ); |
| 2891 | } |
| 2892 | |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 2893 | void YUY2ToUV422Row_SSE2(const uint8* src_yuy2, |
| 2894 | uint8* dst_u, uint8* dst_v, int pix) { |
| 2895 | asm volatile ( |
| 2896 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2897 | "psrlw $0x8,%%xmm5 \n" |
| 2898 | "sub %1,%2 \n" |
| 2899 | ".p2align 4 \n" |
| 2900 | "1: \n" |
| 2901 | "movdqa (%0),%%xmm0 \n" |
| 2902 | "movdqa 0x10(%0),%%xmm1 \n" |
| 2903 | "lea 0x20(%0),%0 \n" |
| 2904 | "psrlw $0x8,%%xmm0 \n" |
| 2905 | "psrlw $0x8,%%xmm1 \n" |
| 2906 | "packuswb %%xmm1,%%xmm0 \n" |
| 2907 | "movdqa %%xmm0,%%xmm1 \n" |
| 2908 | "pand %%xmm5,%%xmm0 \n" |
| 2909 | "packuswb %%xmm0,%%xmm0 \n" |
| 2910 | "psrlw $0x8,%%xmm1 \n" |
| 2911 | "packuswb %%xmm1,%%xmm1 \n" |
| 2912 | "movq %%xmm0,(%1) \n" |
| 2913 | "movq %%xmm1,(%1,%2) \n" |
| 2914 | "lea 0x8(%1),%1 \n" |
| 2915 | "sub $0x10,%3 \n" |
| 2916 | "jg 1b \n" |
| 2917 | : "+r"(src_yuy2), // %0 |
| 2918 | "+r"(dst_u), // %1 |
| 2919 | "+r"(dst_v), // %2 |
| 2920 | "+r"(pix) // %3 |
| 2921 | : |
| 2922 | : "memory", "cc" |
| 2923 | #if defined(__SSE2__) |
| 2924 | , "xmm0", "xmm1", "xmm5" |
| 2925 | #endif |
| 2926 | ); |
| 2927 | } |
fbarchard@google.com | f3fb7b6 | 2012-03-29 23:19:34 +0000 | [diff] [blame] | 2928 | |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2929 | void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2, |
| 2930 | uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2931 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2932 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2933 | "psrlw $0x8,%%xmm5 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2934 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2935 | "1: \n" |
| 2936 | "movdqu (%0),%%xmm0 \n" |
| 2937 | "movdqu 0x10(%0),%%xmm1 \n" |
| 2938 | "lea 0x20(%0),%0 \n" |
| 2939 | "pand %%xmm5,%%xmm0 \n" |
| 2940 | "pand %%xmm5,%%xmm1 \n" |
| 2941 | "packuswb %%xmm1,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2942 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2943 | "movdqu %%xmm0,(%1) \n" |
| 2944 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2945 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2946 | : "+r"(src_yuy2), // %0 |
| 2947 | "+r"(dst_y), // %1 |
| 2948 | "+r"(pix) // %2 |
| 2949 | : |
| 2950 | : "memory", "cc" |
| 2951 | #if defined(__SSE2__) |
| 2952 | , "xmm0", "xmm1", "xmm5" |
| 2953 | #endif |
| 2954 | ); |
| 2955 | } |
| 2956 | |
| 2957 | void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, |
| 2958 | int stride_yuy2, |
fbarchard@google.com | d4164fb | 2012-08-30 20:42:01 +0000 | [diff] [blame] | 2959 | uint8* dst_u, uint8* dst_v, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 2960 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2961 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 2962 | "psrlw $0x8,%%xmm5 \n" |
| 2963 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 2964 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2965 | "1: \n" |
| 2966 | "movdqu (%0),%%xmm0 \n" |
| 2967 | "movdqu 0x10(%0),%%xmm1 \n" |
| 2968 | "movdqu (%0,%4,1),%%xmm2 \n" |
| 2969 | "movdqu 0x10(%0,%4,1),%%xmm3 \n" |
| 2970 | "lea 0x20(%0),%0 \n" |
| 2971 | "pavgb %%xmm2,%%xmm0 \n" |
| 2972 | "pavgb %%xmm3,%%xmm1 \n" |
| 2973 | "psrlw $0x8,%%xmm0 \n" |
| 2974 | "psrlw $0x8,%%xmm1 \n" |
| 2975 | "packuswb %%xmm1,%%xmm0 \n" |
| 2976 | "movdqa %%xmm0,%%xmm1 \n" |
| 2977 | "pand %%xmm5,%%xmm0 \n" |
| 2978 | "packuswb %%xmm0,%%xmm0 \n" |
| 2979 | "psrlw $0x8,%%xmm1 \n" |
| 2980 | "packuswb %%xmm1,%%xmm1 \n" |
| 2981 | "movq %%xmm0,(%1) \n" |
| 2982 | "movq %%xmm1,(%1,%2) \n" |
| 2983 | "lea 0x8(%1),%1 \n" |
| 2984 | "sub $0x10,%3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 2985 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2986 | : "+r"(src_yuy2), // %0 |
| 2987 | "+r"(dst_u), // %1 |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 2988 | "+r"(dst_v), // %2 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2989 | "+r"(pix) // %3 |
| 2990 | : "r"(static_cast<intptr_t>(stride_yuy2)) // %4 |
| 2991 | : "memory", "cc" |
| 2992 | #if defined(__SSE2__) |
| 2993 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" |
| 2994 | #endif |
fbarchard@google.com | f1b6063 | 2012-02-17 19:27:20 +0000 | [diff] [blame] | 2995 | ); |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 2996 | } |
| 2997 | |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 2998 | void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2, |
| 2999 | uint8* dst_u, uint8* dst_v, int pix) { |
| 3000 | asm volatile ( |
| 3001 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3002 | "psrlw $0x8,%%xmm5 \n" |
| 3003 | "sub %1,%2 \n" |
| 3004 | ".p2align 4 \n" |
| 3005 | "1: \n" |
| 3006 | "movdqu (%0),%%xmm0 \n" |
| 3007 | "movdqu 0x10(%0),%%xmm1 \n" |
| 3008 | "lea 0x20(%0),%0 \n" |
| 3009 | "psrlw $0x8,%%xmm0 \n" |
| 3010 | "psrlw $0x8,%%xmm1 \n" |
| 3011 | "packuswb %%xmm1,%%xmm0 \n" |
| 3012 | "movdqa %%xmm0,%%xmm1 \n" |
| 3013 | "pand %%xmm5,%%xmm0 \n" |
| 3014 | "packuswb %%xmm0,%%xmm0 \n" |
| 3015 | "psrlw $0x8,%%xmm1 \n" |
| 3016 | "packuswb %%xmm1,%%xmm1 \n" |
| 3017 | "movq %%xmm0,(%1) \n" |
| 3018 | "movq %%xmm1,(%1,%2) \n" |
| 3019 | "lea 0x8(%1),%1 \n" |
| 3020 | "sub $0x10,%3 \n" |
| 3021 | "jg 1b \n" |
| 3022 | : "+r"(src_yuy2), // %0 |
| 3023 | "+r"(dst_u), // %1 |
| 3024 | "+r"(dst_v), // %2 |
| 3025 | "+r"(pix) // %3 |
| 3026 | : |
| 3027 | : "memory", "cc" |
| 3028 | #if defined(__SSE2__) |
| 3029 | , "xmm0", "xmm1", "xmm5" |
| 3030 | #endif |
| 3031 | ); |
| 3032 | } |
| 3033 | |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3034 | void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 3035 | asm volatile ( |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3036 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3037 | "1: \n" |
| 3038 | "movdqa (%0),%%xmm0 \n" |
| 3039 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3040 | "lea 0x20(%0),%0 \n" |
| 3041 | "psrlw $0x8,%%xmm0 \n" |
| 3042 | "psrlw $0x8,%%xmm1 \n" |
| 3043 | "packuswb %%xmm1,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 3044 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3045 | "movdqa %%xmm0,(%1) \n" |
| 3046 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 3047 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3048 | : "+r"(src_uyvy), // %0 |
| 3049 | "+r"(dst_y), // %1 |
| 3050 | "+r"(pix) // %2 |
| 3051 | : |
| 3052 | : "memory", "cc" |
| 3053 | #if defined(__SSE2__) |
| 3054 | , "xmm0", "xmm1" |
| 3055 | #endif |
| 3056 | ); |
| 3057 | } |
| 3058 | |
| 3059 | void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy, |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 3060 | uint8* dst_u, uint8* dst_v, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 3061 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3062 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3063 | "psrlw $0x8,%%xmm5 \n" |
| 3064 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3065 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3066 | "1: \n" |
| 3067 | "movdqa (%0),%%xmm0 \n" |
| 3068 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3069 | "movdqa (%0,%4,1),%%xmm2 \n" |
| 3070 | "movdqa 0x10(%0,%4,1),%%xmm3 \n" |
| 3071 | "lea 0x20(%0),%0 \n" |
| 3072 | "pavgb %%xmm2,%%xmm0 \n" |
| 3073 | "pavgb %%xmm3,%%xmm1 \n" |
| 3074 | "pand %%xmm5,%%xmm0 \n" |
| 3075 | "pand %%xmm5,%%xmm1 \n" |
| 3076 | "packuswb %%xmm1,%%xmm0 \n" |
| 3077 | "movdqa %%xmm0,%%xmm1 \n" |
| 3078 | "pand %%xmm5,%%xmm0 \n" |
| 3079 | "packuswb %%xmm0,%%xmm0 \n" |
| 3080 | "psrlw $0x8,%%xmm1 \n" |
| 3081 | "packuswb %%xmm1,%%xmm1 \n" |
| 3082 | "movq %%xmm0,(%1) \n" |
| 3083 | "movq %%xmm1,(%1,%2) \n" |
| 3084 | "lea 0x8(%1),%1 \n" |
| 3085 | "sub $0x10,%3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 3086 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3087 | : "+r"(src_uyvy), // %0 |
| 3088 | "+r"(dst_u), // %1 |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 3089 | "+r"(dst_v), // %2 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3090 | "+r"(pix) // %3 |
| 3091 | : "r"(static_cast<intptr_t>(stride_uyvy)) // %4 |
| 3092 | : "memory", "cc" |
| 3093 | #if defined(__SSE2__) |
| 3094 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" |
| 3095 | #endif |
| 3096 | ); |
| 3097 | } |
| 3098 | |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 3099 | void UYVYToUV422Row_SSE2(const uint8* src_uyvy, |
| 3100 | uint8* dst_u, uint8* dst_v, int pix) { |
| 3101 | asm volatile ( |
| 3102 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3103 | "psrlw $0x8,%%xmm5 \n" |
| 3104 | "sub %1,%2 \n" |
| 3105 | ".p2align 4 \n" |
| 3106 | "1: \n" |
| 3107 | "movdqa (%0),%%xmm0 \n" |
| 3108 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3109 | "lea 0x20(%0),%0 \n" |
| 3110 | "pand %%xmm5,%%xmm0 \n" |
| 3111 | "pand %%xmm5,%%xmm1 \n" |
| 3112 | "packuswb %%xmm1,%%xmm0 \n" |
| 3113 | "movdqa %%xmm0,%%xmm1 \n" |
| 3114 | "pand %%xmm5,%%xmm0 \n" |
| 3115 | "packuswb %%xmm0,%%xmm0 \n" |
| 3116 | "psrlw $0x8,%%xmm1 \n" |
| 3117 | "packuswb %%xmm1,%%xmm1 \n" |
| 3118 | "movq %%xmm0,(%1) \n" |
| 3119 | "movq %%xmm1,(%1,%2) \n" |
| 3120 | "lea 0x8(%1),%1 \n" |
| 3121 | "sub $0x10,%3 \n" |
| 3122 | "jg 1b \n" |
| 3123 | : "+r"(src_uyvy), // %0 |
| 3124 | "+r"(dst_u), // %1 |
| 3125 | "+r"(dst_v), // %2 |
| 3126 | "+r"(pix) // %3 |
| 3127 | : |
| 3128 | : "memory", "cc" |
| 3129 | #if defined(__SSE2__) |
| 3130 | , "xmm0", "xmm1", "xmm5" |
| 3131 | #endif |
| 3132 | ); |
| 3133 | } |
| 3134 | |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3135 | void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy, |
| 3136 | uint8* dst_y, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 3137 | asm volatile ( |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3138 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3139 | "1: \n" |
| 3140 | "movdqu (%0),%%xmm0 \n" |
| 3141 | "movdqu 0x10(%0),%%xmm1 \n" |
| 3142 | "lea 0x20(%0),%0 \n" |
| 3143 | "psrlw $0x8,%%xmm0 \n" |
| 3144 | "psrlw $0x8,%%xmm1 \n" |
| 3145 | "packuswb %%xmm1,%%xmm0 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 3146 | "sub $0x10,%2 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3147 | "movdqu %%xmm0,(%1) \n" |
| 3148 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 3149 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3150 | : "+r"(src_uyvy), // %0 |
| 3151 | "+r"(dst_y), // %1 |
| 3152 | "+r"(pix) // %2 |
| 3153 | : |
| 3154 | : "memory", "cc" |
| 3155 | #if defined(__SSE2__) |
| 3156 | , "xmm0", "xmm1" |
| 3157 | #endif |
| 3158 | ); |
| 3159 | } |
| 3160 | |
| 3161 | void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy, |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 3162 | uint8* dst_u, uint8* dst_v, int pix) { |
fbarchard@google.com | e14b2ab | 2012-03-26 16:15:15 +0000 | [diff] [blame] | 3163 | asm volatile ( |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3164 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3165 | "psrlw $0x8,%%xmm5 \n" |
| 3166 | "sub %1,%2 \n" |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3167 | ".p2align 4 \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3168 | "1: \n" |
| 3169 | "movdqu (%0),%%xmm0 \n" |
| 3170 | "movdqu 0x10(%0),%%xmm1 \n" |
| 3171 | "movdqu (%0,%4,1),%%xmm2 \n" |
| 3172 | "movdqu 0x10(%0,%4,1),%%xmm3 \n" |
| 3173 | "lea 0x20(%0),%0 \n" |
| 3174 | "pavgb %%xmm2,%%xmm0 \n" |
| 3175 | "pavgb %%xmm3,%%xmm1 \n" |
| 3176 | "pand %%xmm5,%%xmm0 \n" |
| 3177 | "pand %%xmm5,%%xmm1 \n" |
| 3178 | "packuswb %%xmm1,%%xmm0 \n" |
| 3179 | "movdqa %%xmm0,%%xmm1 \n" |
| 3180 | "pand %%xmm5,%%xmm0 \n" |
| 3181 | "packuswb %%xmm0,%%xmm0 \n" |
| 3182 | "psrlw $0x8,%%xmm1 \n" |
| 3183 | "packuswb %%xmm1,%%xmm1 \n" |
| 3184 | "movq %%xmm0,(%1) \n" |
| 3185 | "movq %%xmm1,(%1,%2) \n" |
| 3186 | "lea 0x8(%1),%1 \n" |
| 3187 | "sub $0x10,%3 \n" |
fbarchard@google.com | 18184fd | 2012-03-12 18:53:19 +0000 | [diff] [blame] | 3188 | "jg 1b \n" |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3189 | : "+r"(src_uyvy), // %0 |
| 3190 | "+r"(dst_u), // %1 |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 3191 | "+r"(dst_v), // %2 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3192 | "+r"(pix) // %3 |
| 3193 | : "r"(static_cast<intptr_t>(stride_uyvy)) // %4 |
| 3194 | : "memory", "cc" |
| 3195 | #if defined(__SSE2__) |
| 3196 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" |
| 3197 | #endif |
| 3198 | ); |
| 3199 | } |
fbarchard@google.com | c704f78 | 2012-08-30 19:53:48 +0000 | [diff] [blame] | 3200 | |
| 3201 | void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy, |
| 3202 | uint8* dst_u, uint8* dst_v, int pix) { |
| 3203 | asm volatile ( |
| 3204 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3205 | "psrlw $0x8,%%xmm5 \n" |
| 3206 | "sub %1,%2 \n" |
| 3207 | ".p2align 4 \n" |
| 3208 | "1: \n" |
| 3209 | "movdqu (%0),%%xmm0 \n" |
| 3210 | "movdqu 0x10(%0),%%xmm1 \n" |
| 3211 | "lea 0x20(%0),%0 \n" |
| 3212 | "pand %%xmm5,%%xmm0 \n" |
| 3213 | "pand %%xmm5,%%xmm1 \n" |
| 3214 | "packuswb %%xmm1,%%xmm0 \n" |
| 3215 | "movdqa %%xmm0,%%xmm1 \n" |
| 3216 | "pand %%xmm5,%%xmm0 \n" |
| 3217 | "packuswb %%xmm0,%%xmm0 \n" |
| 3218 | "psrlw $0x8,%%xmm1 \n" |
| 3219 | "packuswb %%xmm1,%%xmm1 \n" |
| 3220 | "movq %%xmm0,(%1) \n" |
| 3221 | "movq %%xmm1,(%1,%2) \n" |
| 3222 | "lea 0x8(%1),%1 \n" |
| 3223 | "sub $0x10,%3 \n" |
| 3224 | "jg 1b \n" |
| 3225 | : "+r"(src_uyvy), // %0 |
| 3226 | "+r"(dst_u), // %1 |
| 3227 | "+r"(dst_v), // %2 |
| 3228 | "+r"(pix) // %3 |
| 3229 | : |
| 3230 | : "memory", "cc" |
| 3231 | #if defined(__SSE2__) |
| 3232 | , "xmm0", "xmm1", "xmm5" |
| 3233 | #endif |
| 3234 | ); |
| 3235 | } |
fbarchard@google.com | b95dbf2 | 2012-02-11 01:18:30 +0000 | [diff] [blame] | 3236 | #endif // HAS_YUY2TOYROW_SSE2 |
fbarchard@google.com | e5f3fd4 | 2012-02-06 22:40:32 +0000 | [diff] [blame] | 3237 | |
fbarchard@google.com | ba3aeed | 2012-03-07 18:24:51 +0000 | [diff] [blame] | 3238 | #ifdef HAS_ARGBBLENDROW_SSE2 |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3239 | // Blend 8 pixels at a time. |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3240 | void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, |
| 3241 | uint8* dst_argb, int width) { |
fbarchard@google.com | c757f30 | 2012-04-03 00:49:16 +0000 | [diff] [blame] | 3242 | asm volatile ( |
| 3243 | "pcmpeqb %%xmm7,%%xmm7 \n" |
| 3244 | "psrlw $0xf,%%xmm7 \n" |
| 3245 | "pcmpeqb %%xmm6,%%xmm6 \n" |
| 3246 | "psrlw $0x8,%%xmm6 \n" |
| 3247 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3248 | "psllw $0x8,%%xmm5 \n" |
| 3249 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 3250 | "pslld $0x18,%%xmm4 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3251 | "sub $0x1,%3 \n" |
| 3252 | "je 91f \n" |
| 3253 | "jl 99f \n" |
fbarchard@google.com | c757f30 | 2012-04-03 00:49:16 +0000 | [diff] [blame] | 3254 | |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3255 | // 1 pixel loop until destination pointer is aligned. |
| 3256 | "10: \n" |
| 3257 | "test $0xf,%2 \n" |
| 3258 | "je 19f \n" |
fbarchard@google.com | c757f30 | 2012-04-03 00:49:16 +0000 | [diff] [blame] | 3259 | "movd (%0),%%xmm3 \n" |
| 3260 | "lea 0x4(%0),%0 \n" |
| 3261 | "movdqa %%xmm3,%%xmm0 \n" |
| 3262 | "pxor %%xmm4,%%xmm3 \n" |
| 3263 | "movd (%1),%%xmm2 \n" |
| 3264 | "psrlw $0x8,%%xmm3 \n" |
| 3265 | "pshufhw $0xf5,%%xmm3,%%xmm3 \n" |
| 3266 | "pshuflw $0xf5,%%xmm3,%%xmm3 \n" |
| 3267 | "pand %%xmm6,%%xmm2 \n" |
| 3268 | "paddw %%xmm7,%%xmm3 \n" |
| 3269 | "pmullw %%xmm3,%%xmm2 \n" |
| 3270 | "movd (%1),%%xmm1 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3271 | "lea 0x4(%1),%1 \n" |
fbarchard@google.com | c757f30 | 2012-04-03 00:49:16 +0000 | [diff] [blame] | 3272 | "psrlw $0x8,%%xmm1 \n" |
| 3273 | "por %%xmm4,%%xmm0 \n" |
| 3274 | "pmullw %%xmm3,%%xmm1 \n" |
| 3275 | "psrlw $0x8,%%xmm2 \n" |
| 3276 | "paddusb %%xmm2,%%xmm0 \n" |
| 3277 | "pand %%xmm5,%%xmm1 \n" |
| 3278 | "paddusb %%xmm1,%%xmm0 \n" |
| 3279 | "sub $0x1,%3 \n" |
| 3280 | "movd %%xmm0,(%2) \n" |
| 3281 | "lea 0x4(%2),%2 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3282 | "jge 10b \n" |
| 3283 | |
| 3284 | "19: \n" |
fbarchard@google.com | ee22088 | 2012-06-14 00:07:56 +0000 | [diff] [blame] | 3285 | "add $1-4,%3 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3286 | "jl 49f \n" |
| 3287 | |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3288 | // 4 pixel loop. |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3289 | ".p2align 2 \n" |
| 3290 | "41: \n" |
| 3291 | "movdqu (%0),%%xmm3 \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3292 | "lea 0x10(%0),%0 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3293 | "movdqa %%xmm3,%%xmm0 \n" |
| 3294 | "pxor %%xmm4,%%xmm3 \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3295 | "movdqu (%1),%%xmm2 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3296 | "psrlw $0x8,%%xmm3 \n" |
| 3297 | "pshufhw $0xf5,%%xmm3,%%xmm3 \n" |
| 3298 | "pshuflw $0xf5,%%xmm3,%%xmm3 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3299 | "pand %%xmm6,%%xmm2 \n" |
| 3300 | "paddw %%xmm7,%%xmm3 \n" |
| 3301 | "pmullw %%xmm3,%%xmm2 \n" |
| 3302 | "movdqu (%1),%%xmm1 \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3303 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3304 | "psrlw $0x8,%%xmm1 \n" |
| 3305 | "por %%xmm4,%%xmm0 \n" |
| 3306 | "pmullw %%xmm3,%%xmm1 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3307 | "psrlw $0x8,%%xmm2 \n" |
| 3308 | "paddusb %%xmm2,%%xmm0 \n" |
| 3309 | "pand %%xmm5,%%xmm1 \n" |
| 3310 | "paddusb %%xmm1,%%xmm0 \n" |
| 3311 | "sub $0x4,%3 \n" |
| 3312 | "movdqa %%xmm0,(%2) \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3313 | "lea 0x10(%2),%2 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3314 | "jge 41b \n" |
| 3315 | |
| 3316 | "49: \n" |
| 3317 | "add $0x3,%3 \n" |
| 3318 | "jl 99f \n" |
| 3319 | |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3320 | // 1 pixel loop. |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3321 | "91: \n" |
| 3322 | "movd (%0),%%xmm3 \n" |
| 3323 | "lea 0x4(%0),%0 \n" |
| 3324 | "movdqa %%xmm3,%%xmm0 \n" |
| 3325 | "pxor %%xmm4,%%xmm3 \n" |
| 3326 | "movd (%1),%%xmm2 \n" |
| 3327 | "psrlw $0x8,%%xmm3 \n" |
| 3328 | "pshufhw $0xf5,%%xmm3,%%xmm3 \n" |
| 3329 | "pshuflw $0xf5,%%xmm3,%%xmm3 \n" |
| 3330 | "pand %%xmm6,%%xmm2 \n" |
| 3331 | "paddw %%xmm7,%%xmm3 \n" |
| 3332 | "pmullw %%xmm3,%%xmm2 \n" |
| 3333 | "movd (%1),%%xmm1 \n" |
| 3334 | "lea 0x4(%1),%1 \n" |
| 3335 | "psrlw $0x8,%%xmm1 \n" |
| 3336 | "por %%xmm4,%%xmm0 \n" |
| 3337 | "pmullw %%xmm3,%%xmm1 \n" |
| 3338 | "psrlw $0x8,%%xmm2 \n" |
| 3339 | "paddusb %%xmm2,%%xmm0 \n" |
| 3340 | "pand %%xmm5,%%xmm1 \n" |
| 3341 | "paddusb %%xmm1,%%xmm0 \n" |
| 3342 | "sub $0x1,%3 \n" |
| 3343 | "movd %%xmm0,(%2) \n" |
| 3344 | "lea 0x4(%2),%2 \n" |
| 3345 | "jge 91b \n" |
| 3346 | "99: \n" |
| 3347 | : "+r"(src_argb0), // %0 |
| 3348 | "+r"(src_argb1), // %1 |
| 3349 | "+r"(dst_argb), // %2 |
| 3350 | "+r"(width) // %3 |
fbarchard@google.com | c757f30 | 2012-04-03 00:49:16 +0000 | [diff] [blame] | 3351 | : |
| 3352 | : "memory", "cc" |
| 3353 | #if defined(__SSE2__) |
| 3354 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
| 3355 | #endif |
| 3356 | ); |
| 3357 | } |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3358 | #endif // HAS_ARGBBLENDROW_SSE2 |
fbarchard@google.com | c757f30 | 2012-04-03 00:49:16 +0000 | [diff] [blame] | 3359 | |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3360 | #ifdef HAS_ARGBBLENDROW_SSSE3 |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3361 | // Shuffle table for isolating alpha. |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3362 | CONST uvec8 kShuffleAlpha = { |
| 3363 | 3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80, |
| 3364 | 11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80 |
| 3365 | }; |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3366 | |
| 3367 | // Blend 8 pixels at a time |
| 3368 | // Shuffle table for reversing the bytes. |
| 3369 | |
| 3370 | // Same as SSE2, but replaces |
| 3371 | // psrlw xmm3, 8 // alpha |
| 3372 | // pshufhw xmm3, xmm3,0F5h // 8 alpha words |
| 3373 | // pshuflw xmm3, xmm3,0F5h |
| 3374 | // with.. |
| 3375 | // pshufb xmm3, kShuffleAlpha // alpha |
| 3376 | |
| 3377 | void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1, |
| 3378 | uint8* dst_argb, int width) { |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3379 | asm volatile ( |
| 3380 | "pcmpeqb %%xmm7,%%xmm7 \n" |
| 3381 | "psrlw $0xf,%%xmm7 \n" |
| 3382 | "pcmpeqb %%xmm6,%%xmm6 \n" |
| 3383 | "psrlw $0x8,%%xmm6 \n" |
| 3384 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3385 | "psllw $0x8,%%xmm5 \n" |
| 3386 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 3387 | "pslld $0x18,%%xmm4 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3388 | "sub $0x1,%3 \n" |
| 3389 | "je 91f \n" |
| 3390 | "jl 99f \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3391 | |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3392 | // 1 pixel loop until destination pointer is aligned. |
| 3393 | "10: \n" |
| 3394 | "test $0xf,%2 \n" |
| 3395 | "je 19f \n" |
| 3396 | "movd (%0),%%xmm3 \n" |
| 3397 | "lea 0x4(%0),%0 \n" |
| 3398 | "movdqa %%xmm3,%%xmm0 \n" |
| 3399 | "pxor %%xmm4,%%xmm3 \n" |
| 3400 | "movd (%1),%%xmm2 \n" |
| 3401 | "pshufb %4,%%xmm3 \n" |
| 3402 | "pand %%xmm6,%%xmm2 \n" |
| 3403 | "paddw %%xmm7,%%xmm3 \n" |
| 3404 | "pmullw %%xmm3,%%xmm2 \n" |
| 3405 | "movd (%1),%%xmm1 \n" |
| 3406 | "lea 0x4(%1),%1 \n" |
| 3407 | "psrlw $0x8,%%xmm1 \n" |
| 3408 | "por %%xmm4,%%xmm0 \n" |
| 3409 | "pmullw %%xmm3,%%xmm1 \n" |
| 3410 | "psrlw $0x8,%%xmm2 \n" |
| 3411 | "paddusb %%xmm2,%%xmm0 \n" |
| 3412 | "pand %%xmm5,%%xmm1 \n" |
| 3413 | "paddusb %%xmm1,%%xmm0 \n" |
| 3414 | "sub $0x1,%3 \n" |
| 3415 | "movd %%xmm0,(%2) \n" |
| 3416 | "lea 0x4(%2),%2 \n" |
| 3417 | "jge 10b \n" |
| 3418 | |
| 3419 | "19: \n" |
fbarchard@google.com | ee22088 | 2012-06-14 00:07:56 +0000 | [diff] [blame] | 3420 | "add $1-4,%3 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3421 | "jl 49f \n" |
fbarchard@google.com | f877e71 | 2012-08-15 00:51:24 +0000 | [diff] [blame] | 3422 | "test $0xf,%0 \n" |
| 3423 | "jne 41f \n" |
| 3424 | "test $0xf,%1 \n" |
| 3425 | "jne 41f \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3426 | |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3427 | // 4 pixel loop. |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3428 | ".p2align 2 \n" |
fbarchard@google.com | f877e71 | 2012-08-15 00:51:24 +0000 | [diff] [blame] | 3429 | "40: \n" |
fbarchard@google.com | f3181b3 | 2012-08-15 18:57:18 +0000 | [diff] [blame] | 3430 | "movdqa (%0),%%xmm3 \n" |
fbarchard@google.com | f877e71 | 2012-08-15 00:51:24 +0000 | [diff] [blame] | 3431 | "lea 0x10(%0),%0 \n" |
| 3432 | "movdqa %%xmm3,%%xmm0 \n" |
| 3433 | "pxor %%xmm4,%%xmm3 \n" |
fbarchard@google.com | f3181b3 | 2012-08-15 18:57:18 +0000 | [diff] [blame] | 3434 | "movdqa (%1),%%xmm2 \n" |
fbarchard@google.com | f877e71 | 2012-08-15 00:51:24 +0000 | [diff] [blame] | 3435 | "pshufb %4,%%xmm3 \n" |
| 3436 | "pand %%xmm6,%%xmm2 \n" |
| 3437 | "paddw %%xmm7,%%xmm3 \n" |
| 3438 | "pmullw %%xmm3,%%xmm2 \n" |
fbarchard@google.com | f3181b3 | 2012-08-15 18:57:18 +0000 | [diff] [blame] | 3439 | "movdqa (%1),%%xmm1 \n" |
fbarchard@google.com | f877e71 | 2012-08-15 00:51:24 +0000 | [diff] [blame] | 3440 | "lea 0x10(%1),%1 \n" |
| 3441 | "psrlw $0x8,%%xmm1 \n" |
| 3442 | "por %%xmm4,%%xmm0 \n" |
| 3443 | "pmullw %%xmm3,%%xmm1 \n" |
| 3444 | "psrlw $0x8,%%xmm2 \n" |
| 3445 | "paddusb %%xmm2,%%xmm0 \n" |
| 3446 | "pand %%xmm5,%%xmm1 \n" |
| 3447 | "paddusb %%xmm1,%%xmm0 \n" |
| 3448 | "sub $0x4,%3 \n" |
| 3449 | "movdqa %%xmm0,(%2) \n" |
| 3450 | "lea 0x10(%2),%2 \n" |
| 3451 | "jge 40b \n" |
| 3452 | "jmp 49f \n" |
| 3453 | |
| 3454 | // 4 pixel unaligned loop. |
| 3455 | ".p2align 2 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3456 | "41: \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3457 | "movdqu (%0),%%xmm3 \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3458 | "lea 0x10(%0),%0 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3459 | "movdqa %%xmm3,%%xmm0 \n" |
| 3460 | "pxor %%xmm4,%%xmm3 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3461 | "movdqu (%1),%%xmm2 \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3462 | "pshufb %4,%%xmm3 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3463 | "pand %%xmm6,%%xmm2 \n" |
| 3464 | "paddw %%xmm7,%%xmm3 \n" |
| 3465 | "pmullw %%xmm3,%%xmm2 \n" |
| 3466 | "movdqu (%1),%%xmm1 \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3467 | "lea 0x10(%1),%1 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3468 | "psrlw $0x8,%%xmm1 \n" |
| 3469 | "por %%xmm4,%%xmm0 \n" |
| 3470 | "pmullw %%xmm3,%%xmm1 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3471 | "psrlw $0x8,%%xmm2 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3472 | "paddusb %%xmm2,%%xmm0 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3473 | "pand %%xmm5,%%xmm1 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3474 | "paddusb %%xmm1,%%xmm0 \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3475 | "sub $0x4,%3 \n" |
| 3476 | "movdqa %%xmm0,(%2) \n" |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3477 | "lea 0x10(%2),%2 \n" |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3478 | "jge 41b \n" |
| 3479 | |
| 3480 | "49: \n" |
| 3481 | "add $0x3,%3 \n" |
| 3482 | "jl 99f \n" |
| 3483 | |
fbarchard@google.com | 794fe12 | 2012-06-15 01:05:01 +0000 | [diff] [blame] | 3484 | // 1 pixel loop. |
fbarchard@google.com | bac5f2c | 2012-06-13 23:43:12 +0000 | [diff] [blame] | 3485 | "91: \n" |
| 3486 | "movd (%0),%%xmm3 \n" |
| 3487 | "lea 0x4(%0),%0 \n" |
| 3488 | "movdqa %%xmm3,%%xmm0 \n" |
| 3489 | "pxor %%xmm4,%%xmm3 \n" |
| 3490 | "movd (%1),%%xmm2 \n" |
| 3491 | "pshufb %4,%%xmm3 \n" |
| 3492 | "pand %%xmm6,%%xmm2 \n" |
| 3493 | "paddw %%xmm7,%%xmm3 \n" |
| 3494 | "pmullw %%xmm3,%%xmm2 \n" |
| 3495 | "movd (%1),%%xmm1 \n" |
| 3496 | "lea 0x4(%1),%1 \n" |
| 3497 | "psrlw $0x8,%%xmm1 \n" |
| 3498 | "por %%xmm4,%%xmm0 \n" |
| 3499 | "pmullw %%xmm3,%%xmm1 \n" |
| 3500 | "psrlw $0x8,%%xmm2 \n" |
| 3501 | "paddusb %%xmm2,%%xmm0 \n" |
| 3502 | "pand %%xmm5,%%xmm1 \n" |
| 3503 | "paddusb %%xmm1,%%xmm0 \n" |
| 3504 | "sub $0x1,%3 \n" |
| 3505 | "movd %%xmm0,(%2) \n" |
| 3506 | "lea 0x4(%2),%2 \n" |
| 3507 | "jge 91b \n" |
| 3508 | "99: \n" |
fbarchard@google.com | 96af870 | 2012-04-06 18:22:27 +0000 | [diff] [blame] | 3509 | : "+r"(src_argb0), // %0 |
| 3510 | "+r"(src_argb1), // %1 |
| 3511 | "+r"(dst_argb), // %2 |
| 3512 | "+r"(width) // %3 |
| 3513 | : "m"(kShuffleAlpha) // %4 |
| 3514 | : "memory", "cc" |
| 3515 | #if defined(__SSE2__) |
| 3516 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
| 3517 | #endif |
| 3518 | ); |
| 3519 | } |
| 3520 | #endif // HAS_ARGBBLENDROW_SSSE3 |
| 3521 | |
fbarchard@google.com | 8ed5422 | 2012-04-18 17:07:07 +0000 | [diff] [blame] | 3522 | #ifdef HAS_ARGBATTENUATE_SSE2 |
| 3523 | // Attenuate 4 pixels at a time. |
| 3524 | // aligned to 16 bytes |
| 3525 | void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) { |
| 3526 | asm volatile ( |
| 3527 | "sub %0,%1 \n" |
| 3528 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 3529 | "pslld $0x18,%%xmm4 \n" |
| 3530 | "pcmpeqb %%xmm5,%%xmm5 \n" |
| 3531 | "psrld $0x8,%%xmm5 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3532 | |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3533 | // 4 pixel loop. |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3534 | ".p2align 4 \n" |
fbarchard@google.com | 8ed5422 | 2012-04-18 17:07:07 +0000 | [diff] [blame] | 3535 | "1: \n" |
| 3536 | "movdqa (%0),%%xmm0 \n" |
| 3537 | "punpcklbw %%xmm0,%%xmm0 \n" |
| 3538 | "pshufhw $0xff,%%xmm0,%%xmm2 \n" |
| 3539 | "pshuflw $0xff,%%xmm2,%%xmm2 \n" |
| 3540 | "pmulhuw %%xmm2,%%xmm0 \n" |
| 3541 | "movdqa (%0),%%xmm1 \n" |
| 3542 | "punpckhbw %%xmm1,%%xmm1 \n" |
| 3543 | "pshufhw $0xff,%%xmm1,%%xmm2 \n" |
| 3544 | "pshuflw $0xff,%%xmm2,%%xmm2 \n" |
| 3545 | "pmulhuw %%xmm2,%%xmm1 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3546 | "movdqa (%0),%%xmm2 \n" |
fbarchard@google.com | 8ed5422 | 2012-04-18 17:07:07 +0000 | [diff] [blame] | 3547 | "psrlw $0x8,%%xmm0 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3548 | "pand %%xmm4,%%xmm2 \n" |
fbarchard@google.com | 8ed5422 | 2012-04-18 17:07:07 +0000 | [diff] [blame] | 3549 | "psrlw $0x8,%%xmm1 \n" |
| 3550 | "packuswb %%xmm1,%%xmm0 \n" |
| 3551 | "pand %%xmm5,%%xmm0 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3552 | "por %%xmm2,%%xmm0 \n" |
fbarchard@google.com | 8ed5422 | 2012-04-18 17:07:07 +0000 | [diff] [blame] | 3553 | "sub $0x4,%2 \n" |
| 3554 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 3555 | "lea 0x10(%0),%0 \n" |
| 3556 | "jg 1b \n" |
| 3557 | : "+r"(src_argb), // %0 |
| 3558 | "+r"(dst_argb), // %1 |
| 3559 | "+r"(width) // %2 |
| 3560 | : |
| 3561 | : "memory", "cc" |
| 3562 | #if defined(__SSE2__) |
| 3563 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 3564 | #endif |
| 3565 | ); |
| 3566 | } |
| 3567 | #endif // HAS_ARGBATTENUATE_SSE2 |
| 3568 | |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3569 | #ifdef HAS_ARGBATTENUATEROW_SSSE3 |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3570 | // Shuffle table duplicating alpha |
| 3571 | CONST uvec8 kShuffleAlpha0 = { |
| 3572 | 3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u, |
| 3573 | }; |
| 3574 | CONST uvec8 kShuffleAlpha1 = { |
| 3575 | 11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u, |
| 3576 | 15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u, |
| 3577 | }; |
| 3578 | // Attenuate 4 pixels at a time. |
| 3579 | // aligned to 16 bytes |
| 3580 | void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) { |
| 3581 | asm volatile ( |
| 3582 | "sub %0,%1 \n" |
| 3583 | "pcmpeqb %%xmm3,%%xmm3 \n" |
| 3584 | "pslld $0x18,%%xmm3 \n" |
| 3585 | "movdqa %3,%%xmm4 \n" |
| 3586 | "movdqa %4,%%xmm5 \n" |
| 3587 | |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3588 | // 4 pixel loop. |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3589 | ".p2align 4 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3590 | "1: \n" |
| 3591 | "movdqa (%0),%%xmm0 \n" |
| 3592 | "pshufb %%xmm4,%%xmm0 \n" |
| 3593 | "movdqa (%0),%%xmm1 \n" |
| 3594 | "punpcklbw %%xmm1,%%xmm1 \n" |
| 3595 | "pmulhuw %%xmm1,%%xmm0 \n" |
| 3596 | "movdqa (%0),%%xmm1 \n" |
| 3597 | "pshufb %%xmm5,%%xmm1 \n" |
| 3598 | "movdqa (%0),%%xmm2 \n" |
| 3599 | "punpckhbw %%xmm2,%%xmm2 \n" |
| 3600 | "pmulhuw %%xmm2,%%xmm1 \n" |
| 3601 | "movdqa (%0),%%xmm2 \n" |
| 3602 | "pand %%xmm3,%%xmm2 \n" |
| 3603 | "psrlw $0x8,%%xmm0 \n" |
| 3604 | "psrlw $0x8,%%xmm1 \n" |
| 3605 | "packuswb %%xmm1,%%xmm0 \n" |
| 3606 | "por %%xmm2,%%xmm0 \n" |
| 3607 | "sub $0x4,%2 \n" |
| 3608 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 3609 | "lea 0x10(%0),%0 \n" |
| 3610 | "jg 1b \n" |
| 3611 | : "+r"(src_argb), // %0 |
| 3612 | "+r"(dst_argb), // %1 |
| 3613 | "+r"(width) // %2 |
| 3614 | : "m"(kShuffleAlpha0), // %3 |
| 3615 | "m"(kShuffleAlpha1) // %4 |
| 3616 | : "memory", "cc" |
| 3617 | #if defined(__SSE2__) |
| 3618 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 3619 | #endif |
| 3620 | ); |
| 3621 | } |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3622 | #endif // HAS_ARGBATTENUATEROW_SSSE3 |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3623 | |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3624 | #ifdef HAS_ARGBUNATTENUATEROW_SSE2 |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3625 | // Unattenuate 4 pixels at a time. |
| 3626 | // aligned to 16 bytes |
| 3627 | void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, |
| 3628 | int width) { |
| 3629 | uintptr_t alpha = 0; |
| 3630 | asm volatile ( |
| 3631 | "sub %0,%1 \n" |
| 3632 | "pcmpeqb %%xmm4,%%xmm4 \n" |
| 3633 | "pslld $0x18,%%xmm4 \n" |
| 3634 | |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3635 | // 4 pixel loop. |
fbarchard@google.com | 5bf29b5 | 2012-05-02 00:10:16 +0000 | [diff] [blame] | 3636 | ".p2align 4 \n" |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3637 | "1: \n" |
| 3638 | "movdqa (%0),%%xmm0 \n" |
| 3639 | "movzb 0x3(%0),%3 \n" |
| 3640 | "punpcklbw %%xmm0,%%xmm0 \n" |
| 3641 | "movd 0x0(%4,%3,4),%%xmm2 \n" |
| 3642 | "movzb 0x7(%0),%3 \n" |
| 3643 | "movd 0x0(%4,%3,4),%%xmm3 \n" |
| 3644 | "pshuflw $0xc0,%%xmm2,%%xmm2 \n" |
| 3645 | "pshuflw $0xc0,%%xmm3,%%xmm3 \n" |
| 3646 | "movlhps %%xmm3,%%xmm2 \n" |
| 3647 | "pmulhuw %%xmm2,%%xmm0 \n" |
| 3648 | "movdqa (%0),%%xmm1 \n" |
| 3649 | "movzb 0xb(%0),%3 \n" |
| 3650 | "punpckhbw %%xmm1,%%xmm1 \n" |
| 3651 | "movd 0x0(%4,%3,4),%%xmm2 \n" |
| 3652 | "movzb 0xf(%0),%3 \n" |
| 3653 | "movd 0x0(%4,%3,4),%%xmm3 \n" |
| 3654 | "pshuflw $0xc0,%%xmm2,%%xmm2 \n" |
| 3655 | "pshuflw $0xc0,%%xmm3,%%xmm3 \n" |
| 3656 | "movlhps %%xmm3,%%xmm2 \n" |
| 3657 | "pmulhuw %%xmm2,%%xmm1 \n" |
| 3658 | "movdqa (%0),%%xmm2 \n" |
| 3659 | "pand %%xmm4,%%xmm2 \n" |
| 3660 | "packuswb %%xmm1,%%xmm0 \n" |
| 3661 | "por %%xmm2,%%xmm0 \n" |
| 3662 | "sub $0x4,%2 \n" |
| 3663 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 3664 | "lea 0x10(%0),%0 \n" |
| 3665 | "jg 1b \n" |
| 3666 | : "+r"(src_argb), // %0 |
| 3667 | "+r"(dst_argb), // %1 |
| 3668 | "+r"(width), // %2 |
| 3669 | "+r"(alpha) // %3 |
| 3670 | : "r"(fixed_invtbl8) // %4 |
| 3671 | : "memory", "cc" |
| 3672 | #if defined(__SSE2__) |
| 3673 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 3674 | #endif |
| 3675 | ); |
| 3676 | } |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3677 | #endif // HAS_ARGBUNATTENUATEROW_SSE2 |
fbarchard@google.com | 810cd91 | 2012-04-20 20:15:27 +0000 | [diff] [blame] | 3678 | |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3679 | #ifdef HAS_ARGBGRAYROW_SSSE3 |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 3680 | // Constant for ARGB color to gray scale. 0.11 * B + 0.59 * G + 0.30 * R |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3681 | CONST vec8 kARGBToGray = { |
| 3682 | 14, 76, 38, 0, 14, 76, 38, 0, 14, 76, 38, 0, 14, 76, 38, 0 |
| 3683 | }; |
| 3684 | |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3685 | // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3686 | void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) { |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3687 | asm volatile ( |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3688 | "movdqa %3,%%xmm4 \n" |
| 3689 | "sub %0,%1 \n" |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3690 | |
| 3691 | // 8 pixel loop. |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3692 | ".p2align 4 \n" |
| 3693 | "1: \n" |
| 3694 | "movdqa (%0),%%xmm0 \n" |
| 3695 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3696 | "pmaddubsw %%xmm4,%%xmm0 \n" |
| 3697 | "pmaddubsw %%xmm4,%%xmm1 \n" |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3698 | "phaddw %%xmm1,%%xmm0 \n" |
| 3699 | "psrlw $0x7,%%xmm0 \n" |
| 3700 | "packuswb %%xmm0,%%xmm0 \n" |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3701 | "movdqa (%0),%%xmm2 \n" |
| 3702 | "movdqa 0x10(%0),%%xmm3 \n" |
| 3703 | "psrld $0x18,%%xmm2 \n" |
| 3704 | "psrld $0x18,%%xmm3 \n" |
| 3705 | "packuswb %%xmm3,%%xmm2 \n" |
| 3706 | "packuswb %%xmm2,%%xmm2 \n" |
| 3707 | "movdqa %%xmm0,%%xmm3 \n" |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3708 | "punpcklbw %%xmm0,%%xmm0 \n" |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3709 | "punpcklbw %%xmm2,%%xmm3 \n" |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3710 | "movdqa %%xmm0,%%xmm1 \n" |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3711 | "punpcklwd %%xmm3,%%xmm0 \n" |
| 3712 | "punpckhwd %%xmm3,%%xmm1 \n" |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3713 | "sub $0x8,%2 \n" |
| 3714 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 3715 | "movdqa %%xmm1,0x10(%0,%1,1) \n" |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3716 | "lea 0x20(%0),%0 \n" |
| 3717 | "jg 1b \n" |
fbarchard@google.com | eeac290 | 2012-07-18 18:54:32 +0000 | [diff] [blame] | 3718 | : "+r"(src_argb), // %0 |
| 3719 | "+r"(dst_argb), // %1 |
| 3720 | "+r"(width) // %2 |
| 3721 | : "m"(kARGBToGray) // %3 |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3722 | : "memory", "cc" |
| 3723 | #if defined(__SSE2__) |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3724 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" |
fbarchard@google.com | ffaea7e | 2012-05-18 19:43:59 +0000 | [diff] [blame] | 3725 | #endif |
| 3726 | ); |
| 3727 | } |
| 3728 | #endif // HAS_ARGBGRAYROW_SSSE3 |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3729 | |
| 3730 | #ifdef HAS_ARGBSEPIAROW_SSSE3 |
| 3731 | // b = (r * 35 + g * 68 + b * 17) >> 7 |
| 3732 | // g = (r * 45 + g * 88 + b * 22) >> 7 |
| 3733 | // r = (r * 50 + g * 98 + b * 24) >> 7 |
| 3734 | // Constant for ARGB color to sepia tone |
| 3735 | CONST vec8 kARGBToSepiaB = { |
| 3736 | 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0 |
| 3737 | }; |
| 3738 | |
| 3739 | CONST vec8 kARGBToSepiaG = { |
| 3740 | 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0 |
| 3741 | }; |
| 3742 | |
| 3743 | CONST vec8 kARGBToSepiaR = { |
| 3744 | 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0 |
| 3745 | }; |
| 3746 | |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3747 | // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3748 | void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) { |
| 3749 | asm volatile ( |
| 3750 | "movdqa %2,%%xmm2 \n" |
| 3751 | "movdqa %3,%%xmm3 \n" |
| 3752 | "movdqa %4,%%xmm4 \n" |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3753 | |
| 3754 | // 8 pixel loop. |
fbarchard@google.com | 221e602 | 2012-05-21 22:24:41 +0000 | [diff] [blame] | 3755 | ".p2align 4 \n" |
| 3756 | "1: \n" |
| 3757 | "movdqa (%0),%%xmm0 \n" |
| 3758 | "movdqa 0x10(%0),%%xmm6 \n" |
| 3759 | "pmaddubsw %%xmm2,%%xmm0 \n" |
| 3760 | "pmaddubsw %%xmm2,%%xmm6 \n" |
| 3761 | "phaddw %%xmm6,%%xmm0 \n" |
| 3762 | "psrlw $0x7,%%xmm0 \n" |
| 3763 | "packuswb %%xmm0,%%xmm0 \n" |
| 3764 | "movdqa (%0),%%xmm5 \n" |
| 3765 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3766 | "pmaddubsw %%xmm3,%%xmm5 \n" |
| 3767 | "pmaddubsw %%xmm3,%%xmm1 \n" |
| 3768 | "phaddw %%xmm1,%%xmm5 \n" |
| 3769 | "psrlw $0x7,%%xmm5 \n" |
| 3770 | "packuswb %%xmm5,%%xmm5 \n" |
| 3771 | "punpcklbw %%xmm5,%%xmm0 \n" |
| 3772 | "movdqa (%0),%%xmm5 \n" |
| 3773 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3774 | "pmaddubsw %%xmm4,%%xmm5 \n" |
| 3775 | "pmaddubsw %%xmm4,%%xmm1 \n" |
| 3776 | "phaddw %%xmm1,%%xmm5 \n" |
| 3777 | "psrlw $0x7,%%xmm5 \n" |
| 3778 | "packuswb %%xmm5,%%xmm5 \n" |
| 3779 | "movdqa (%0),%%xmm6 \n" |
| 3780 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3781 | "psrld $0x18,%%xmm6 \n" |
| 3782 | "psrld $0x18,%%xmm1 \n" |
| 3783 | "packuswb %%xmm1,%%xmm6 \n" |
| 3784 | "packuswb %%xmm6,%%xmm6 \n" |
| 3785 | "punpcklbw %%xmm6,%%xmm5 \n" |
| 3786 | "movdqa %%xmm0,%%xmm1 \n" |
| 3787 | "punpcklwd %%xmm5,%%xmm0 \n" |
| 3788 | "punpckhwd %%xmm5,%%xmm1 \n" |
| 3789 | "sub $0x8,%1 \n" |
| 3790 | "movdqa %%xmm0,(%0) \n" |
| 3791 | "movdqa %%xmm1,0x10(%0) \n" |
| 3792 | "lea 0x20(%0),%0 \n" |
| 3793 | "jg 1b \n" |
| 3794 | : "+r"(dst_argb), // %0 |
| 3795 | "+r"(width) // %1 |
| 3796 | : "m"(kARGBToSepiaB), // %2 |
| 3797 | "m"(kARGBToSepiaG), // %3 |
| 3798 | "m"(kARGBToSepiaR) // %4 |
| 3799 | : "memory", "cc" |
| 3800 | #if defined(__SSE2__) |
| 3801 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
| 3802 | #endif |
| 3803 | ); |
| 3804 | } |
| 3805 | #endif // HAS_ARGBSEPIAROW_SSSE3 |
| 3806 | |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3807 | #ifdef HAS_ARGBCOLORMATRIXROW_SSSE3 |
| 3808 | // Tranform 8 ARGB pixels (32 bytes) with color matrix. |
| 3809 | // Same as Sepia except matrix is provided. |
| 3810 | void ARGBColorMatrixRow_SSSE3(uint8* dst_argb, const int8* matrix_argb, |
| 3811 | int width) { |
| 3812 | asm volatile ( |
| 3813 | "movd (%2),%%xmm2 \n" |
| 3814 | "movd 0x4(%2),%%xmm3 \n" |
| 3815 | "movd 0x8(%2),%%xmm4 \n" |
| 3816 | "pshufd $0x0,%%xmm2,%%xmm2 \n" |
| 3817 | "pshufd $0x0,%%xmm3,%%xmm3 \n" |
| 3818 | "pshufd $0x0,%%xmm4,%%xmm4 \n" |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3819 | |
| 3820 | // 8 pixel loop. |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3821 | ".p2align 4 \n" |
| 3822 | "1: \n" |
| 3823 | "movdqa (%0),%%xmm0 \n" |
| 3824 | "movdqa 0x10(%0),%%xmm6 \n" |
| 3825 | "pmaddubsw %%xmm2,%%xmm0 \n" |
| 3826 | "pmaddubsw %%xmm2,%%xmm6 \n" |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3827 | "movdqa (%0),%%xmm5 \n" |
| 3828 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3829 | "pmaddubsw %%xmm3,%%xmm5 \n" |
| 3830 | "pmaddubsw %%xmm3,%%xmm1 \n" |
fbarchard@google.com | 8f439ea | 2012-06-21 02:00:34 +0000 | [diff] [blame] | 3831 | "phaddsw %%xmm6,%%xmm0 \n" |
| 3832 | "phaddsw %%xmm1,%%xmm5 \n" |
| 3833 | "psraw $0x7,%%xmm0 \n" |
| 3834 | "psraw $0x7,%%xmm5 \n" |
| 3835 | "packuswb %%xmm0,%%xmm0 \n" |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3836 | "packuswb %%xmm5,%%xmm5 \n" |
| 3837 | "punpcklbw %%xmm5,%%xmm0 \n" |
| 3838 | "movdqa (%0),%%xmm5 \n" |
| 3839 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3840 | "pmaddubsw %%xmm4,%%xmm5 \n" |
| 3841 | "pmaddubsw %%xmm4,%%xmm1 \n" |
fbarchard@google.com | 8f439ea | 2012-06-21 02:00:34 +0000 | [diff] [blame] | 3842 | "phaddsw %%xmm1,%%xmm5 \n" |
| 3843 | "psraw $0x7,%%xmm5 \n" |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3844 | "packuswb %%xmm5,%%xmm5 \n" |
| 3845 | "movdqa (%0),%%xmm6 \n" |
| 3846 | "movdqa 0x10(%0),%%xmm1 \n" |
| 3847 | "psrld $0x18,%%xmm6 \n" |
| 3848 | "psrld $0x18,%%xmm1 \n" |
| 3849 | "packuswb %%xmm1,%%xmm6 \n" |
| 3850 | "packuswb %%xmm6,%%xmm6 \n" |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3851 | "movdqa %%xmm0,%%xmm1 \n" |
fbarchard@google.com | 8f439ea | 2012-06-21 02:00:34 +0000 | [diff] [blame] | 3852 | "punpcklbw %%xmm6,%%xmm5 \n" |
fbarchard@google.com | e442dc4 | 2012-06-18 17:37:09 +0000 | [diff] [blame] | 3853 | "punpcklwd %%xmm5,%%xmm0 \n" |
| 3854 | "punpckhwd %%xmm5,%%xmm1 \n" |
| 3855 | "sub $0x8,%1 \n" |
| 3856 | "movdqa %%xmm0,(%0) \n" |
| 3857 | "movdqa %%xmm1,0x10(%0) \n" |
| 3858 | "lea 0x20(%0),%0 \n" |
| 3859 | "jg 1b \n" |
| 3860 | : "+r"(dst_argb), // %0 |
| 3861 | "+r"(width) // %1 |
| 3862 | : "r"(matrix_argb) // %2 |
| 3863 | : "memory", "cc" |
| 3864 | #if defined(__SSE2__) |
| 3865 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
| 3866 | #endif |
| 3867 | ); |
| 3868 | } |
| 3869 | #endif // HAS_ARGBCOLORMATRIXROW_SSSE3 |
| 3870 | |
fbarchard@google.com | 81b804e | 2012-06-20 02:15:01 +0000 | [diff] [blame] | 3871 | #ifdef HAS_ARGBQUANTIZEROW_SSE2 |
| 3872 | // Quantize 4 ARGB pixels (16 bytes). |
| 3873 | // aligned to 16 bytes |
| 3874 | void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size, |
| 3875 | int interval_offset, int width) { |
| 3876 | asm volatile ( |
| 3877 | "movd %2,%%xmm2 \n" |
| 3878 | "movd %3,%%xmm3 \n" |
| 3879 | "movd %4,%%xmm4 \n" |
| 3880 | "pshuflw $0x40,%%xmm2,%%xmm2 \n" |
| 3881 | "pshufd $0x44,%%xmm2,%%xmm2 \n" |
| 3882 | "pshuflw $0x40,%%xmm3,%%xmm3 \n" |
| 3883 | "pshufd $0x44,%%xmm3,%%xmm3 \n" |
| 3884 | "pshuflw $0x40,%%xmm4,%%xmm4 \n" |
| 3885 | "pshufd $0x44,%%xmm4,%%xmm4 \n" |
| 3886 | "pxor %%xmm5,%%xmm5 \n" |
| 3887 | "pcmpeqb %%xmm6,%%xmm6 \n" |
| 3888 | "pslld $0x18,%%xmm6 \n" |
| 3889 | |
| 3890 | // 4 pixel loop. |
| 3891 | ".p2align 2 \n" |
| 3892 | "1: \n" |
| 3893 | "movdqa (%0),%%xmm0 \n" |
| 3894 | "punpcklbw %%xmm5,%%xmm0 \n" |
| 3895 | "pmulhuw %%xmm2,%%xmm0 \n" |
| 3896 | "movdqa (%0),%%xmm1 \n" |
| 3897 | "punpckhbw %%xmm5,%%xmm1 \n" |
| 3898 | "pmulhuw %%xmm2,%%xmm1 \n" |
| 3899 | "pmullw %%xmm3,%%xmm0 \n" |
| 3900 | "movdqa (%0),%%xmm7 \n" |
| 3901 | "pmullw %%xmm3,%%xmm1 \n" |
| 3902 | "pand %%xmm6,%%xmm7 \n" |
| 3903 | "paddw %%xmm4,%%xmm0 \n" |
| 3904 | "paddw %%xmm4,%%xmm1 \n" |
| 3905 | "packuswb %%xmm1,%%xmm0 \n" |
| 3906 | "por %%xmm7,%%xmm0 \n" |
| 3907 | "sub $0x4,%1 \n" |
| 3908 | "movdqa %%xmm0,(%0) \n" |
| 3909 | "lea 0x10(%0),%0 \n" |
| 3910 | "jg 1b \n" |
| 3911 | : "+r"(dst_argb), // %0 |
| 3912 | "+r"(width) // %1 |
| 3913 | : "r"(scale), // %2 |
| 3914 | "r"(interval_size), // %3 |
| 3915 | "r"(interval_offset) // %4 |
| 3916 | : "memory", "cc" |
| 3917 | #if defined(__SSE2__) |
| 3918 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
| 3919 | #endif |
| 3920 | ); |
| 3921 | } |
| 3922 | #endif // HAS_ARGBQUANTIZEROW_SSE2 |
| 3923 | |
fbarchard@google.com | f51e879 | 2012-06-10 02:40:04 +0000 | [diff] [blame] | 3924 | #ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2 |
| 3925 | // Creates a table of cumulative sums where each value is a sum of all values |
| 3926 | // above and to the left of the value, inclusive of the value. |
| 3927 | void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum, |
fbarchard@google.com | 133adc4 | 2012-06-12 05:46:49 +0000 | [diff] [blame] | 3928 | const int32* previous_cumsum, int width) { |
fbarchard@google.com | f51e879 | 2012-06-10 02:40:04 +0000 | [diff] [blame] | 3929 | asm volatile ( |
| 3930 | "sub %1,%2 \n" |
| 3931 | "pxor %%xmm0,%%xmm0 \n" |
| 3932 | "pxor %%xmm1,%%xmm1 \n" |
| 3933 | "sub $0x4,%3 \n" |
| 3934 | "jl 49f \n" |
| 3935 | "test $0xf,%1 \n" |
| 3936 | "jne 49f \n" |
| 3937 | |
| 3938 | // 4 pixel loop \n" |
| 3939 | ".p2align 2 \n" |
| 3940 | "40: \n" |
| 3941 | "movdqu (%0),%%xmm2 \n" |
| 3942 | "lea 0x10(%0),%0 \n" |
| 3943 | "movdqa %%xmm2,%%xmm4 \n" |
| 3944 | "punpcklbw %%xmm1,%%xmm2 \n" |
| 3945 | "movdqa %%xmm2,%%xmm3 \n" |
| 3946 | "punpcklwd %%xmm1,%%xmm2 \n" |
| 3947 | "punpckhwd %%xmm1,%%xmm3 \n" |
| 3948 | "punpckhbw %%xmm1,%%xmm4 \n" |
| 3949 | "movdqa %%xmm4,%%xmm5 \n" |
| 3950 | "punpcklwd %%xmm1,%%xmm4 \n" |
| 3951 | "punpckhwd %%xmm1,%%xmm5 \n" |
| 3952 | "paddd %%xmm2,%%xmm0 \n" |
| 3953 | "movdqa (%1,%2,1),%%xmm2 \n" |
| 3954 | "paddd %%xmm0,%%xmm2 \n" |
| 3955 | "paddd %%xmm3,%%xmm0 \n" |
| 3956 | "movdqa 0x10(%1,%2,1),%%xmm3 \n" |
| 3957 | "paddd %%xmm0,%%xmm3 \n" |
| 3958 | "paddd %%xmm4,%%xmm0 \n" |
| 3959 | "movdqa 0x20(%1,%2,1),%%xmm4 \n" |
| 3960 | "paddd %%xmm0,%%xmm4 \n" |
| 3961 | "paddd %%xmm5,%%xmm0 \n" |
| 3962 | "movdqa 0x30(%1,%2,1),%%xmm5 \n" |
| 3963 | "paddd %%xmm0,%%xmm5 \n" |
| 3964 | "movdqa %%xmm2,(%1) \n" |
| 3965 | "movdqa %%xmm3,0x10(%1) \n" |
| 3966 | "movdqa %%xmm4,0x20(%1) \n" |
| 3967 | "movdqa %%xmm5,0x30(%1) \n" |
| 3968 | "lea 0x40(%1),%1 \n" |
| 3969 | "sub $0x4,%3 \n" |
| 3970 | "jge 40b \n" |
| 3971 | |
| 3972 | "49: \n" |
| 3973 | "add $0x3,%3 \n" |
| 3974 | "jl 19f \n" |
| 3975 | |
| 3976 | // 1 pixel loop \n" |
| 3977 | ".p2align 2 \n" |
| 3978 | "10: \n" |
| 3979 | "movd (%0),%%xmm2 \n" |
| 3980 | "lea 0x4(%0),%0 \n" |
fbarchard@google.com | f38aefe | 2012-06-12 00:29:29 +0000 | [diff] [blame] | 3981 | "punpcklbw %%xmm1,%%xmm2 \n" |
| 3982 | "punpcklwd %%xmm1,%%xmm2 \n" |
fbarchard@google.com | f51e879 | 2012-06-10 02:40:04 +0000 | [diff] [blame] | 3983 | "paddd %%xmm2,%%xmm0 \n" |
| 3984 | "movdqu (%1,%2,1),%%xmm2 \n" |
| 3985 | "paddd %%xmm0,%%xmm2 \n" |
| 3986 | "movdqu %%xmm2,(%1) \n" |
| 3987 | "lea 0x10(%1),%1 \n" |
| 3988 | "sub $0x1,%3 \n" |
| 3989 | "jge 10b \n" |
| 3990 | |
| 3991 | "19: \n" |
| 3992 | : "+r"(row), // %0 |
| 3993 | "+r"(cumsum), // %1 |
| 3994 | "+r"(previous_cumsum), // %2 |
| 3995 | "+r"(width) // %3 |
| 3996 | : |
| 3997 | : "memory", "cc" |
| 3998 | #if defined(__SSE2__) |
| 3999 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
| 4000 | #endif |
| 4001 | ); |
| 4002 | } |
| 4003 | #endif // HAS_COMPUTECUMULATIVESUMROW_SSE2 |
| 4004 | |
| 4005 | #ifdef HAS_CUMULATIVESUMTOAVERAGE_SSE2 |
| 4006 | void CumulativeSumToAverage_SSE2(const int32* topleft, const int32* botleft, |
| 4007 | int width, int area, uint8* dst, int count) { |
| 4008 | asm volatile ( |
| 4009 | "movd %5,%%xmm4 \n" |
| 4010 | "cvtdq2ps %%xmm4,%%xmm4 \n" |
| 4011 | "rcpss %%xmm4,%%xmm4 \n" |
| 4012 | "pshufd $0x0,%%xmm4,%%xmm4 \n" |
| 4013 | "sub $0x4,%3 \n" |
| 4014 | "jl 49f \n" |
| 4015 | |
| 4016 | // 4 pixel loop \n" |
| 4017 | ".p2align 2 \n" |
| 4018 | "40: \n" |
| 4019 | "movdqa (%0),%%xmm0 \n" |
| 4020 | "movdqa 0x10(%0),%%xmm1 \n" |
| 4021 | "movdqa 0x20(%0),%%xmm2 \n" |
| 4022 | "movdqa 0x30(%0),%%xmm3 \n" |
| 4023 | "psubd (%0,%4,4),%%xmm0 \n" |
| 4024 | "psubd 0x10(%0,%4,4),%%xmm1 \n" |
| 4025 | "psubd 0x20(%0,%4,4),%%xmm2 \n" |
| 4026 | "psubd 0x30(%0,%4,4),%%xmm3 \n" |
| 4027 | "lea 0x40(%0),%0 \n" |
| 4028 | "psubd (%1),%%xmm0 \n" |
| 4029 | "psubd 0x10(%1),%%xmm1 \n" |
| 4030 | "psubd 0x20(%1),%%xmm2 \n" |
| 4031 | "psubd 0x30(%1),%%xmm3 \n" |
| 4032 | "paddd (%1,%4,4),%%xmm0 \n" |
| 4033 | "paddd 0x10(%1,%4,4),%%xmm1 \n" |
| 4034 | "paddd 0x20(%1,%4,4),%%xmm2 \n" |
| 4035 | "paddd 0x30(%1,%4,4),%%xmm3 \n" |
| 4036 | "lea 0x40(%1),%1 \n" |
| 4037 | "cvtdq2ps %%xmm0,%%xmm0 \n" |
| 4038 | "cvtdq2ps %%xmm1,%%xmm1 \n" |
| 4039 | "mulps %%xmm4,%%xmm0 \n" |
| 4040 | "mulps %%xmm4,%%xmm1 \n" |
| 4041 | "cvtdq2ps %%xmm2,%%xmm2 \n" |
| 4042 | "cvtdq2ps %%xmm3,%%xmm3 \n" |
| 4043 | "mulps %%xmm4,%%xmm2 \n" |
| 4044 | "mulps %%xmm4,%%xmm3 \n" |
| 4045 | "cvtps2dq %%xmm0,%%xmm0 \n" |
| 4046 | "cvtps2dq %%xmm1,%%xmm1 \n" |
| 4047 | "cvtps2dq %%xmm2,%%xmm2 \n" |
| 4048 | "cvtps2dq %%xmm3,%%xmm3 \n" |
| 4049 | "packssdw %%xmm1,%%xmm0 \n" |
| 4050 | "packssdw %%xmm3,%%xmm2 \n" |
| 4051 | "packuswb %%xmm2,%%xmm0 \n" |
| 4052 | "movdqu %%xmm0,(%2) \n" |
| 4053 | "lea 0x10(%2),%2 \n" |
| 4054 | "sub $0x4,%3 \n" |
| 4055 | "jge 40b \n" |
| 4056 | |
| 4057 | "49: \n" |
| 4058 | "add $0x3,%3 \n" |
| 4059 | "jl 19f \n" |
| 4060 | |
| 4061 | // 1 pixel loop \n" |
| 4062 | ".p2align 2 \n" |
| 4063 | "10: \n" |
| 4064 | "movdqa (%0),%%xmm0 \n" |
| 4065 | "psubd (%0,%4,4),%%xmm0 \n" |
| 4066 | "lea 0x10(%0),%0 \n" |
| 4067 | "psubd (%1),%%xmm0 \n" |
| 4068 | "paddd (%1,%4,4),%%xmm0 \n" |
| 4069 | "lea 0x10(%1),%1 \n" |
| 4070 | "cvtdq2ps %%xmm0,%%xmm0 \n" |
| 4071 | "mulps %%xmm4,%%xmm0 \n" |
| 4072 | "cvtps2dq %%xmm0,%%xmm0 \n" |
| 4073 | "packssdw %%xmm0,%%xmm0 \n" |
| 4074 | "packuswb %%xmm0,%%xmm0 \n" |
| 4075 | "movd %%xmm0,(%2) \n" |
| 4076 | "lea 0x4(%2),%2 \n" |
| 4077 | "sub $0x1,%3 \n" |
| 4078 | "jge 10b \n" |
| 4079 | "19: \n" |
| 4080 | : "+r"(topleft), // %0 |
| 4081 | "+r"(botleft), // %1 |
| 4082 | "+r"(dst), // %2 |
| 4083 | "+rm"(count) // %3 |
| 4084 | : "r"(static_cast<intptr_t>(width)), // %4 |
| 4085 | "rm"(area) // %5 |
| 4086 | : "memory", "cc" |
| 4087 | #if defined(__SSE2__) |
| 4088 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" |
| 4089 | #endif |
| 4090 | ); |
| 4091 | } |
| 4092 | #endif // HAS_CUMULATIVESUMTOAVERAGE_SSE2 |
fbarchard@google.com | 6398e1d | 2012-07-11 19:12:32 +0000 | [diff] [blame] | 4093 | #ifdef HAS_ARGBSHADE_SSE2 |
| 4094 | // Shade 4 pixels at a time by specified value. |
| 4095 | // Aligned to 16 bytes. |
| 4096 | void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width, |
| 4097 | uint32 value) { |
| 4098 | asm volatile ( |
| 4099 | "movd %3,%%xmm2 \n" |
| 4100 | "sub %0,%1 \n" |
| 4101 | "punpcklbw %%xmm2,%%xmm2 \n" |
| 4102 | "punpcklqdq %%xmm2,%%xmm2 \n" |
fbarchard@google.com | f51e879 | 2012-06-10 02:40:04 +0000 | [diff] [blame] | 4103 | |
fbarchard@google.com | 6398e1d | 2012-07-11 19:12:32 +0000 | [diff] [blame] | 4104 | // 4 pixel loop. |
| 4105 | ".p2align 2 \n" |
| 4106 | "1: \n" |
| 4107 | "movdqa (%0),%%xmm0 \n" |
| 4108 | "movdqa %%xmm0,%%xmm1 \n" |
| 4109 | "punpcklbw %%xmm0,%%xmm0 \n" |
| 4110 | "punpckhbw %%xmm1,%%xmm1 \n" |
| 4111 | "pmulhuw %%xmm2,%%xmm0 \n" |
| 4112 | "pmulhuw %%xmm2,%%xmm1 \n" |
| 4113 | "psrlw $0x8,%%xmm0 \n" |
| 4114 | "psrlw $0x8,%%xmm1 \n" |
| 4115 | "packuswb %%xmm1,%%xmm0 \n" |
| 4116 | "sub $0x4,%2 \n" |
| 4117 | "movdqa %%xmm0,(%0,%1,1) \n" |
| 4118 | "lea 0x10(%0),%0 \n" |
| 4119 | "jg 1b \n" |
| 4120 | : "+r"(src_argb), // %0 |
| 4121 | "+r"(dst_argb), // %1 |
| 4122 | "+r"(width) // %2 |
| 4123 | : "r"(value) // %3 |
| 4124 | : "memory", "cc" |
| 4125 | #if defined(__SSE2__) |
| 4126 | , "xmm0", "xmm1", "xmm2" |
| 4127 | #endif |
| 4128 | ); |
| 4129 | } |
| 4130 | #endif // HAS_ARGBSHADE_SSE2 |
fbarchard@google.com | f51e879 | 2012-06-10 02:40:04 +0000 | [diff] [blame] | 4131 | |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4132 | #ifdef HAS_ARGBAFFINEROW_SSE2 |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4133 | // TODO(fbarchard): Find 64 bit way to avoid masking. |
| 4134 | // TODO(fbarchard): Investigate why 4 pixels is slower than 2 on Core2. |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4135 | // Copy ARGB pixels from source image with slope to a row of destination. |
fbarchard@google.com | 41f24bf | 2012-08-17 16:51:25 +0000 | [diff] [blame] | 4136 | // Caveat - in 64 bit, movd is used with 64 bit gpr due to Mac gcc producing |
fbarchard@google.com | 64ce0ab | 2012-10-09 00:05:29 +0000 | [diff] [blame] | 4137 | // an error if movq is used. movd %%xmm0,%1 |
fbarchard@google.com | 41f24bf | 2012-08-17 16:51:25 +0000 | [diff] [blame] | 4138 | |
fbarchard@google.com | fc7314e | 2012-09-27 02:17:51 +0000 | [diff] [blame] | 4139 | LIBYUV_API |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4140 | void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride, |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 4141 | uint8* dst_argb, const float* src_dudv, int width) { |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4142 | intptr_t src_argb_stride_temp = src_argb_stride; |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4143 | intptr_t temp = 0; |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4144 | asm volatile ( |
| 4145 | "movq (%3),%%xmm2 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4146 | "movq 0x8(%3),%%xmm7 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4147 | "shl $0x10,%1 \n" |
| 4148 | "add $0x4,%1 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4149 | "movd %1,%%xmm5 \n" |
| 4150 | "sub $0x4,%4 \n" |
| 4151 | "jl 49f \n" |
| 4152 | |
| 4153 | "pshufd $0x44,%%xmm7,%%xmm7 \n" |
| 4154 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4155 | "movdqa %%xmm2,%%xmm0 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4156 | "addps %%xmm7,%%xmm0 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4157 | "movlhps %%xmm0,%%xmm2 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4158 | "movdqa %%xmm7,%%xmm4 \n" |
| 4159 | "addps %%xmm4,%%xmm4 \n" |
| 4160 | "movdqa %%xmm2,%%xmm3 \n" |
| 4161 | "addps %%xmm4,%%xmm3 \n" |
| 4162 | "addps %%xmm4,%%xmm4 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4163 | |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4164 | // 4 pixel loop \n" |
| 4165 | ".p2align 4 \n" |
| 4166 | "40: \n" |
| 4167 | "cvttps2dq %%xmm2,%%xmm0 \n" |
| 4168 | "cvttps2dq %%xmm3,%%xmm1 \n" |
| 4169 | "packssdw %%xmm1,%%xmm0 \n" |
| 4170 | "pmaddwd %%xmm5,%%xmm0 \n" |
| 4171 | #if defined(__x86_64__) |
fbarchard@google.com | 41f24bf | 2012-08-17 16:51:25 +0000 | [diff] [blame] | 4172 | "movd %%xmm0,%1 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4173 | "mov %1,%5 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4174 | "and $0x0fffffff,%1 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4175 | "shr $32,%5 \n" |
| 4176 | "pshufd $0xEE,%%xmm0,%%xmm0 \n" |
| 4177 | #else |
| 4178 | "movd %%xmm0,%1 \n" |
| 4179 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
| 4180 | "movd %%xmm0,%5 \n" |
| 4181 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
| 4182 | #endif |
| 4183 | "movd (%0,%1,1),%%xmm1 \n" |
| 4184 | "movd (%0,%5,1),%%xmm6 \n" |
| 4185 | "punpckldq %%xmm6,%%xmm1 \n" |
| 4186 | "addps %%xmm4,%%xmm2 \n" |
| 4187 | "movq %%xmm1,(%2) \n" |
| 4188 | #if defined(__x86_64__) |
fbarchard@google.com | 41f24bf | 2012-08-17 16:51:25 +0000 | [diff] [blame] | 4189 | "movd %%xmm0,%1 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4190 | "mov %1,%5 \n" |
| 4191 | "and $0x0fffffff,%1 \n" |
| 4192 | "shr $32,%5 \n" |
| 4193 | #else |
| 4194 | "movd %%xmm0,%1 \n" |
| 4195 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
| 4196 | "movd %%xmm0,%5 \n" |
| 4197 | #endif |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4198 | "movd (%0,%1,1),%%xmm0 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4199 | "movd (%0,%5,1),%%xmm6 \n" |
| 4200 | "punpckldq %%xmm6,%%xmm0 \n" |
| 4201 | "addps %%xmm4,%%xmm3 \n" |
| 4202 | "sub $0x4,%4 \n" |
| 4203 | "movq %%xmm0,0x08(%2) \n" |
| 4204 | "lea 0x10(%2),%2 \n" |
| 4205 | "jge 40b \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4206 | |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4207 | "49: \n" |
| 4208 | "add $0x3,%4 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4209 | "jl 19f \n" |
| 4210 | |
| 4211 | // 1 pixel loop \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4212 | ".p2align 4 \n" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4213 | "10: \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4214 | "cvttps2dq %%xmm2,%%xmm0 \n" |
| 4215 | "packssdw %%xmm0,%%xmm0 \n" |
| 4216 | "pmaddwd %%xmm5,%%xmm0 \n" |
| 4217 | "addps %%xmm7,%%xmm2 \n" |
| 4218 | "movd %%xmm0,%1 \n" |
| 4219 | #if defined(__x86_64__) |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4220 | "and $0x0fffffff,%1 \n" |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4221 | #endif |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4222 | "movd (%0,%1,1),%%xmm0 \n" |
| 4223 | "sub $0x1,%4 \n" |
| 4224 | "movd %%xmm0,(%2) \n" |
| 4225 | "lea 0x4(%2),%2 \n" |
| 4226 | "jge 10b \n" |
| 4227 | "19: \n" |
| 4228 | : "+r"(src_argb), // %0 |
| 4229 | "+r"(src_argb_stride_temp), // %1 |
| 4230 | "+r"(dst_argb), // %2 |
fbarchard@google.com | c1f17f1 | 2012-11-06 02:38:45 +0000 | [diff] [blame^] | 4231 | "+r"(src_dudv), // %3 |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4232 | "+rm"(width), // %4 |
| 4233 | "+r"(temp) // %5 |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4234 | : |
| 4235 | : "memory", "cc" |
| 4236 | #if defined(__SSE2__) |
fbarchard@google.com | e3cc769 | 2012-08-10 20:41:27 +0000 | [diff] [blame] | 4237 | , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
fbarchard@google.com | 7344440 | 2012-08-09 17:33:29 +0000 | [diff] [blame] | 4238 | #endif |
| 4239 | ); |
| 4240 | } |
| 4241 | #endif // HAS_ARGBAFFINEROW_SSE2 |
| 4242 | |
fbarchard@google.com | 9bcc9a2 | 2012-09-16 05:00:37 +0000 | [diff] [blame] | 4243 | // Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version |
| 4244 | void ARGBInterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr, |
| 4245 | ptrdiff_t src_stride, int dst_width, |
| 4246 | int source_y_fraction) { |
| 4247 | asm volatile ( |
| 4248 | "sub %1,%0 \n" |
| 4249 | "shr %3 \n" |
| 4250 | "cmp $0x0,%3 \n" |
| 4251 | "je 2f \n" |
| 4252 | "cmp $0x40,%3 \n" |
| 4253 | "je 3f \n" |
| 4254 | "movd %3,%%xmm0 \n" |
| 4255 | "neg %3 \n" |
| 4256 | "add $0x80,%3 \n" |
| 4257 | "movd %3,%%xmm5 \n" |
| 4258 | "punpcklbw %%xmm0,%%xmm5 \n" |
| 4259 | "punpcklwd %%xmm5,%%xmm5 \n" |
| 4260 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
| 4261 | ".p2align 4 \n" |
| 4262 | "1: \n" |
| 4263 | "movdqa (%1),%%xmm0 \n" |
| 4264 | "movdqa (%1,%4,1),%%xmm2 \n" |
| 4265 | "movdqa %%xmm0,%%xmm1 \n" |
| 4266 | "punpcklbw %%xmm2,%%xmm0 \n" |
| 4267 | "punpckhbw %%xmm2,%%xmm1 \n" |
| 4268 | "pmaddubsw %%xmm5,%%xmm0 \n" |
| 4269 | "pmaddubsw %%xmm5,%%xmm1 \n" |
| 4270 | "psrlw $0x7,%%xmm0 \n" |
| 4271 | "psrlw $0x7,%%xmm1 \n" |
| 4272 | "packuswb %%xmm1,%%xmm0 \n" |
| 4273 | "sub $0x4,%2 \n" |
| 4274 | "movdqa %%xmm0,(%1,%0,1) \n" |
| 4275 | "lea 0x10(%1),%1 \n" |
| 4276 | "jg 1b \n" |
| 4277 | "jmp 4f \n" |
| 4278 | ".p2align 4 \n" |
| 4279 | "2: \n" |
| 4280 | "movdqa (%1),%%xmm0 \n" |
| 4281 | "sub $0x4,%2 \n" |
| 4282 | "movdqa %%xmm0,(%1,%0,1) \n" |
| 4283 | "lea 0x10(%1),%1 \n" |
| 4284 | "jg 2b \n" |
| 4285 | "jmp 4f \n" |
| 4286 | ".p2align 4 \n" |
| 4287 | "3: \n" |
| 4288 | "movdqa (%1),%%xmm0 \n" |
| 4289 | "pavgb (%1,%4,1),%%xmm0 \n" |
| 4290 | "sub $0x4,%2 \n" |
| 4291 | "movdqa %%xmm0,(%1,%0,1) \n" |
| 4292 | "lea 0x10(%1),%1 \n" |
| 4293 | "jg 3b \n" |
| 4294 | "4: \n" |
| 4295 | ".p2align 4 \n" |
| 4296 | : "+r"(dst_ptr), // %0 |
| 4297 | "+r"(src_ptr), // %1 |
| 4298 | "+r"(dst_width), // %2 |
| 4299 | "+r"(source_y_fraction) // %3 |
| 4300 | : "r"(static_cast<intptr_t>(src_stride)) // %4 |
| 4301 | : "memory", "cc" |
| 4302 | #if defined(__SSE2__) |
| 4303 | , "xmm0", "xmm1", "xmm2", "xmm5" |
| 4304 | #endif |
| 4305 | ); |
| 4306 | } |
| 4307 | |
fbarchard@google.com | e91bdac | 2012-10-09 21:09:33 +0000 | [diff] [blame] | 4308 | void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride, |
| 4309 | uint8* dst_uv, int pix) { |
| 4310 | asm volatile ( |
| 4311 | "sub %0,%1 \n" |
| 4312 | ".p2align 4 \n" |
| 4313 | "1: \n" |
| 4314 | "movdqa (%0),%%xmm0 \n" |
| 4315 | "pavgb (%0,%3),%%xmm0 \n" |
| 4316 | "sub $0x10,%2 \n" |
| 4317 | "movdqa %%xmm0,(%0,%1) \n" |
| 4318 | "lea 0x10(%0),%0 \n" |
| 4319 | "jg 1b \n" |
| 4320 | : "+r"(src_uv), // %0 |
| 4321 | "+r"(dst_uv), // %1 |
| 4322 | "+r"(pix) // %2 |
| 4323 | : "r"(static_cast<intptr_t>(src_uv_stride)) // %3 |
| 4324 | : "memory", "cc" |
| 4325 | #if defined(__SSE2__) |
| 4326 | , "xmm0" |
| 4327 | #endif |
| 4328 | ); |
| 4329 | } |
fbarchard@google.com | 8d37dd5 | 2012-10-11 00:07:30 +0000 | [diff] [blame] | 4330 | |
| 4331 | void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer, |
| 4332 | uint32 selector, int pix) { |
| 4333 | asm volatile ( |
| 4334 | "movd %3,%%xmm5 \n" |
| 4335 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
| 4336 | ".p2align 4 \n" |
| 4337 | "1: \n" |
| 4338 | "movdqa (%0),%%xmm0 \n" |
| 4339 | "lea 0x10(%0),%0 \n" |
| 4340 | "pshufb %%xmm5,%%xmm0 \n" |
| 4341 | "sub $0x4,%2 \n" |
| 4342 | "movd %%xmm0,(%1) \n" |
| 4343 | "lea 0x4(%1),%1 \n" |
| 4344 | "jg 1b \n" |
| 4345 | : "+r"(src_argb), // %0 |
| 4346 | "+r"(dst_bayer), // %1 |
| 4347 | "+r"(pix) // %2 |
| 4348 | : "g"(selector) // %3 |
| 4349 | : "memory", "cc" |
| 4350 | #if defined(__SSE2__) |
| 4351 | , "xmm0", "xmm5" |
| 4352 | #endif |
| 4353 | ); |
| 4354 | } |
fbarchard@google.com | 9de8867 | 2012-10-12 06:23:33 +0000 | [diff] [blame] | 4355 | |
| 4356 | void I422ToYUY2Row_SSE2(const uint8* src_y, |
| 4357 | const uint8* src_u, |
| 4358 | const uint8* src_v, |
| 4359 | uint8* dst_frame, int width) { |
| 4360 | asm volatile ( |
| 4361 | "sub %1,%2 \n" |
| 4362 | ".p2align 4 \n" |
| 4363 | "1: \n" |
| 4364 | "movq (%1),%%xmm2 \n" |
| 4365 | "movq (%1,%2,1),%%xmm3 \n" |
| 4366 | "lea 0x8(%1),%1 \n" |
| 4367 | "punpcklbw %%xmm3,%%xmm2 \n" |
| 4368 | "movdqa (%0),%%xmm0 \n" |
| 4369 | "lea 0x10(%0),%0 \n" |
| 4370 | "movdqa %%xmm0,%%xmm1 \n" |
| 4371 | "punpcklbw %%xmm2,%%xmm0 \n" |
| 4372 | "punpckhbw %%xmm2,%%xmm1 \n" |
| 4373 | "movdqa %%xmm0,(%3) \n" |
| 4374 | "movdqa %%xmm1,0x10(%3) \n" |
| 4375 | "lea 0x20(%3),%3 \n" |
| 4376 | "sub $0x10,%4 \n" |
| 4377 | "jg 1b \n" |
| 4378 | : "+r"(src_y), // %0 |
| 4379 | "+r"(src_u), // %1 |
| 4380 | "+r"(src_v), // %2 |
| 4381 | "+r"(dst_frame), // %3 |
| 4382 | "+rm"(width) // %4 |
| 4383 | : |
| 4384 | : "memory", "cc" |
| 4385 | #if defined(__SSE2__) |
| 4386 | , "xmm0", "xmm1", "xmm2", "xmm3" |
| 4387 | #endif |
| 4388 | ); |
| 4389 | } |
| 4390 | |
| 4391 | void I422ToUYVYRow_SSE2(const uint8* src_y, |
| 4392 | const uint8* src_u, |
| 4393 | const uint8* src_v, |
| 4394 | uint8* dst_frame, int width) { |
| 4395 | asm volatile ( |
| 4396 | "sub %1,%2 \n" |
| 4397 | ".p2align 4 \n" |
| 4398 | "1: \n" |
| 4399 | "movq (%1),%%xmm2 \n" |
| 4400 | "movq (%1,%2,1),%%xmm3 \n" |
| 4401 | "lea 0x8(%1),%1 \n" |
| 4402 | "punpcklbw %%xmm3,%%xmm2 \n" |
| 4403 | "movdqa (%0),%%xmm0 \n" |
| 4404 | "movdqa %%xmm2,%%xmm1 \n" |
| 4405 | "lea 0x10(%0),%0 \n" |
| 4406 | "punpcklbw %%xmm0,%%xmm1 \n" |
| 4407 | "punpckhbw %%xmm0,%%xmm2 \n" |
| 4408 | "movdqa %%xmm1,(%3) \n" |
| 4409 | "movdqa %%xmm2,0x10(%3) \n" |
| 4410 | "lea 0x20(%3),%3 \n" |
| 4411 | "sub $0x10,%4 \n" |
| 4412 | "jg 1b \n" |
| 4413 | : "+r"(src_y), // %0 |
| 4414 | "+r"(src_u), // %1 |
| 4415 | "+r"(src_v), // %2 |
| 4416 | "+r"(dst_frame), // %3 |
| 4417 | "+rm"(width) // %4 |
| 4418 | : |
| 4419 | : "memory", "cc" |
| 4420 | #if defined(__SSE2__) |
| 4421 | , "xmm0", "xmm1", "xmm2", "xmm3" |
| 4422 | #endif |
| 4423 | ); |
| 4424 | } |
| 4425 | |
fbarchard@google.com | 2d11d43 | 2012-02-16 02:50:39 +0000 | [diff] [blame] | 4426 | #endif // defined(__x86_64__) || defined(__i386__) |
| 4427 | |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 4428 | #ifdef __cplusplus |
mikhal@webrtc.org | aed1cc9 | 2011-09-28 00:06:25 +0000 | [diff] [blame] | 4429 | } // extern "C" |
fbarchard@google.com | fe5ff7e | 2011-12-10 07:45:58 +0000 | [diff] [blame] | 4430 | } // namespace libyuv |
| 4431 | #endif |